test fixes

switch to mark_slow decorator

.travis.yml

@@ -10,5 +10,5 @@ install:
     - docker build . -t baselines-test
 
 script:
-    - flake8 . --show-source --statistics --exclude=baselines/her
+    - flake8 . --show-source --statistics
     - docker run -e RUNSLOW=1 baselines-test pytest -v .

README.md

@@ -98,8 +98,6 @@ python -m baselines.run --alg=deepq --env=PongNoFrameskip-v4 --num_timesteps=1e6
 ```
 
 ## Saving, loading and visualizing models
-
-### Saving and loading the model
 The algorithms serialization API is not properly unified yet; however, there is a simple method to save / restore trained models. 
 `--save_path` and `--load_path` command-line option loads the tensorflow state from a given path before training, and saves it after the training, respectively. 
 Let's imagine you'd like to train ppo2 on Atari Pong,  save the model and then later visualize what has it learnt.
@@ -113,17 +111,8 @@ python -m baselines.run --alg=ppo2 --env=PongNoFrameskip-v4 --num_timesteps=0 --
 
 *NOTE:* Mujoco environments require normalization to work properly, so we wrap them with VecNormalize wrapper. Currently, to ensure the models are saved with normalization (so that trained models can be restored and run without further training) the normalization coefficients are saved as tensorflow variables. This can decrease the performance somewhat, so if you require high-throughput steps with Mujoco and do not need saving/restoring the models, it may make sense to use numpy normalization instead. To do that, set 'use_tf=False` in [baselines/run.py](baselines/run.py#L116). 
 
-### Logging and vizualizing learning curves and other training metrics
-By default, all summary data, including progress, standard output, is saved to a unique directory in a temp folder, specified by a call to Python's [tempfile.gettempdir()](https://docs.python.org/3/library/tempfile.html#tempfile.gettempdir).
-The directory can be changed with the `--log_path` command-line option.
-```bash
-python -m baselines.run --alg=ppo2 --env=PongNoFrameskip-v4 --num_timesteps=2e7 --save_path=~/models/pong_20M_ppo2 --log_path=~/logs/Pong/
-```
-*NOTE:* Please be aware that the logger will overwrite files of the same name in an existing directory, thus it's recommended that folder names be given a unique timestamp to prevent overwritten logs.
-
-Another way the temp directory can be changed is through the use of the `$OPENAI_LOGDIR` environment variable.
-
-For examples on how to load and display the training data, see [here](docs/viz/viz.ipynb).
+## Loading and vizualizing learning curves and other training metrics
+See [here](docs/viz/viz.ipynb) for instructions on how to load and display the training data. 
 
 ## Subpackages
 
@@ -150,7 +139,7 @@ respectively. Note that these results may be not on the latest version of the co
 To cite this repository in publications:
 
     @misc{baselines,
-      author = {Dhariwal, Prafulla and Hesse, Christopher and Klimov, Oleg and Nichol, Alex and Plappert, Matthias and Radford, Alec and Schulman, John and Sidor, Szymon and Tan, Zhenyu and Wu, Yuhuai and Zhokhov, Peter},
+      author = {Dhariwal, Prafulla and Hesse, Christopher and Klimov, Oleg and Nichol, Alex and Plappert, Matthias and Radford, Alec and Schulman, John and Sidor, Szymon and Wu, Yuhuai and Zhokhov, Peter},
       title = {OpenAI Baselines},
       year = {2017},
       publisher = {GitHub},

baselines/a2c/a2c.py

@@ -1,19 +1,22 @@
 import time
+import functools
 import tensorflow as tf
 
 from baselines import logger
 
 from baselines.common import set_global_seeds, explained_variance
-from baselines.common.models import get_network_builder
-from baselines.common.policies import PolicyWithValue
+from baselines.common import tf_util
+from baselines.common.policies import build_policy
 
-from baselines.a2c.utils import InverseLinearTimeDecay
+
+from baselines.a2c.utils import Scheduler, find_trainable_variables
 from baselines.a2c.runner import Runner
 from baselines.ppo2.ppo2 import safemean
-import os.path as osp
 from collections import deque
 
-class Model(tf.keras.Model):
+from tensorflow import losses
+
+class Model(object):
 
     """
     We use this class to :
@@ -27,42 +30,90 @@ class Model(tf.keras.Model):
         save/load():
         - Save load the model
     """
-    def __init__(self, *, ac_space, policy_network, nupdates,
+    def __init__(self, policy, env, nsteps,
             ent_coef=0.01, vf_coef=0.5, max_grad_norm=0.5, lr=7e-4,
-            alpha=0.99, epsilon=1e-5, total_timesteps=int(80e6)):
+            alpha=0.99, epsilon=1e-5, total_timesteps=int(80e6), lrschedule='linear'):
 
-        super(Model, self).__init__(name='A2CModel')
-        self.train_model = PolicyWithValue(ac_space, policy_network, value_network=None, estimate_q=False)
-        lr_schedule = InverseLinearTimeDecay(initial_learning_rate=lr, nupdates=nupdates)
-        self.optimizer = tf.keras.optimizers.RMSprop(learning_rate=lr_schedule, rho=alpha, epsilon=epsilon)
+        sess = tf_util.get_session()
+        nenvs = env.num_envs
+        nbatch = nenvs*nsteps
 
-        self.ent_coef = ent_coef
-        self.vf_coef = vf_coef
-        self.max_grad_norm = max_grad_norm
-        self.step = self.train_model.step
-        self.value = self.train_model.value
-        self.initial_state = self.train_model.initial_state
 
-    @tf.function
-    def train(self, obs, states, rewards, masks, actions, values):
-        advs = rewards - values
-        with tf.GradientTape() as tape:
-            policy_latent = self.train_model.policy_network(obs)
-            pd, _ = self.train_model.pdtype.pdfromlatent(policy_latent)
-            neglogpac = pd.neglogp(actions)
-            entropy = tf.reduce_mean(pd.entropy())
-            vpred = self.train_model.value(obs)
-            vf_loss = tf.reduce_mean(tf.square(vpred - rewards))
-            pg_loss = tf.reduce_mean(advs * neglogpac)
-            loss = pg_loss - entropy * self.ent_coef + vf_loss * self.vf_coef
+        with tf.variable_scope('a2c_model', reuse=tf.AUTO_REUSE):
+            # step_model is used for sampling
+            step_model = policy(nenvs, 1, sess)
 
-        var_list = tape.watched_variables()
-        grads = tape.gradient(loss, var_list)
-        grads, _ = tf.clip_by_global_norm(grads, self.max_grad_norm)
-        grads_and_vars = list(zip(grads, var_list))
-        self.optimizer.apply_gradients(grads_and_vars)
+            # train_model is used to train our network
+            train_model = policy(nbatch, nsteps, sess)
 
-        return pg_loss, vf_loss, entropy
+        A = tf.placeholder(train_model.action.dtype, train_model.action.shape)
+        ADV = tf.placeholder(tf.float32, [nbatch])
+        R = tf.placeholder(tf.float32, [nbatch])
+        LR = tf.placeholder(tf.float32, [])
+
+        # Calculate the loss
+        # Total loss = Policy gradient loss - entropy * entropy coefficient + Value coefficient * value loss
+
+        # Policy loss
+        neglogpac = train_model.pd.neglogp(A)
+        # L = A(s,a) * -logpi(a|s)
+        pg_loss = tf.reduce_mean(ADV * neglogpac)
+
+        # Entropy is used to improve exploration by limiting the premature convergence to suboptimal policy.
+        entropy = tf.reduce_mean(train_model.pd.entropy())
+
+        # Value loss
+        vf_loss = losses.mean_squared_error(tf.squeeze(train_model.vf), R)
+
+        loss = pg_loss - entropy*ent_coef + vf_loss * vf_coef
+
+        # Update parameters using loss
+        # 1. Get the model parameters
+        params = find_trainable_variables("a2c_model")
+
+        # 2. Calculate the gradients
+        grads = tf.gradients(loss, params)
+        if max_grad_norm is not None:
+            # Clip the gradients (normalize)
+            grads, grad_norm = tf.clip_by_global_norm(grads, max_grad_norm)
+        grads = list(zip(grads, params))
+        # zip aggregate each gradient with parameters associated
+        # For instance zip(ABCD, xyza) => Ax, By, Cz, Da
+
+        # 3. Make op for one policy and value update step of A2C
+        trainer = tf.train.RMSPropOptimizer(learning_rate=LR, decay=alpha, epsilon=epsilon)
+
+        _train = trainer.apply_gradients(grads)
+
+        lr = Scheduler(v=lr, nvalues=total_timesteps, schedule=lrschedule)
+
+        def train(obs, states, rewards, masks, actions, values):
+            # Here we calculate advantage A(s,a) = R + yV(s') - V(s)
+            # rewards = R + yV(s')
+            advs = rewards - values
+            for step in range(len(obs)):
+                cur_lr = lr.value()
+
+            td_map = {train_model.X:obs, A:actions, ADV:advs, R:rewards, LR:cur_lr}
+            if states is not None:
+                td_map[train_model.S] = states
+                td_map[train_model.M] = masks
+            policy_loss, value_loss, policy_entropy, _ = sess.run(
+                [pg_loss, vf_loss, entropy, _train],
+                td_map
+            )
+            return policy_loss, value_loss, policy_entropy
+
+
+        self.train = train
+        self.train_model = train_model
+        self.step_model = step_model
+        self.step = step_model.step
+        self.value = step_model.value
+        self.initial_state = step_model.initial_state
+        self.save = functools.partial(tf_util.save_variables, sess=sess)
+        self.load = functools.partial(tf_util.load_variables, sess=sess)
+        tf.global_variables_initializer().run(session=sess)
 
 
 def learn(
@@ -134,53 +185,31 @@ def learn(
 
     set_global_seeds(seed)
 
-    total_timesteps = int(total_timesteps)
-
     # Get the nb of env
     nenvs = env.num_envs
-
-    # Get state_space and action_space
-    ob_space = env.observation_space
-    ac_space = env.action_space
-
-    if isinstance(network, str):
-        network_type = network
-        policy_network_fn = get_network_builder(network_type)(**network_kwargs)
-        policy_network = policy_network_fn(ob_space.shape)
-
-    # Calculate the batch_size
-    nbatch = nenvs * nsteps
-    nupdates = total_timesteps // nbatch
+    policy = build_policy(env, network, **network_kwargs)
 
     # Instantiate the model object (that creates step_model and train_model)
-    model = Model(ac_space=ac_space, policy_network=policy_network, nupdates=nupdates, ent_coef=ent_coef, vf_coef=vf_coef,
-        max_grad_norm=max_grad_norm, lr=lr, alpha=alpha, epsilon=epsilon, total_timesteps=total_timesteps)
-
+    model = Model(policy=policy, env=env, nsteps=nsteps, ent_coef=ent_coef, vf_coef=vf_coef,
+        max_grad_norm=max_grad_norm, lr=lr, alpha=alpha, epsilon=epsilon, total_timesteps=total_timesteps, lrschedule=lrschedule)
     if load_path is not None:
-        load_path = osp.expanduser(load_path)
-        ckpt = tf.train.Checkpoint(model=model)
-        manager = tf.train.CheckpointManager(ckpt, load_path, max_to_keep=None)
-        ckpt.restore(manager.latest_checkpoint)
+        model.load(load_path)
 
     # Instantiate the runner object
     runner = Runner(env, model, nsteps=nsteps, gamma=gamma)
     epinfobuf = deque(maxlen=100)
 
+    # Calculate the batch_size
+    nbatch = nenvs*nsteps
+
     # Start total timer
     tstart = time.time()
 
-    for update in range(1, nupdates+1):
+    for update in range(1, total_timesteps//nbatch+1):
         # Get mini batch of experiences
         obs, states, rewards, masks, actions, values, epinfos = runner.run()
         epinfobuf.extend(epinfos)
 
-        obs = tf.constant(obs)
-        if states is not None:
-            states = tf.constant(states)
-        rewards = tf.constant(rewards)
-        masks = tf.constant(masks)
-        actions = tf.constant(actions)
-        values = tf.constant(values)
         policy_loss, value_loss, policy_entropy = model.train(obs, states, rewards, masks, actions, values)
         nseconds = time.time()-tstart
 

baselines/a2c/runner.py

@@ -1,4 +1,3 @@
-import tensorflow as tf
 import numpy as np
 from baselines.a2c.utils import discount_with_dones
 from baselines.common.runners import AbstractEnvRunner
@@ -16,37 +15,40 @@ class Runner(AbstractEnvRunner):
     def __init__(self, env, model, nsteps=5, gamma=0.99):
         super().__init__(env=env, model=model, nsteps=nsteps)
         self.gamma = gamma
+        self.batch_action_shape = [x if x is not None else -1 for x in model.train_model.action.shape.as_list()]
+        self.ob_dtype = model.train_model.X.dtype.as_numpy_dtype
 
     def run(self):
         # We initialize the lists that will contain the mb of experiences
         mb_obs, mb_rewards, mb_actions, mb_values, mb_dones = [],[],[],[],[]
         mb_states = self.states
         epinfos = []
-        for _ in range(self.nsteps):
+        for n in range(self.nsteps):
             # Given observations, take action and value (V(s))
             # We already have self.obs because Runner superclass run self.obs[:] = env.reset() on init
-            obs = tf.constant(self.obs)
-            actions, values, self.states, _ = self.model.step(obs)
-            actions = actions._numpy()
+            actions, values, states, _ = self.model.step(self.obs, S=self.states, M=self.dones)
+
             # Append the experiences
-            mb_obs.append(self.obs.copy())
+            mb_obs.append(np.copy(self.obs))
             mb_actions.append(actions)
-            mb_values.append(values._numpy())
+            mb_values.append(values)
             mb_dones.append(self.dones)
 
             # Take actions in env and look the results
-            self.obs[:], rewards, self.dones, infos = self.env.step(actions)
+            obs, rewards, dones, infos = self.env.step(actions)
             for info in infos:
                 maybeepinfo = info.get('episode')
                 if maybeepinfo: epinfos.append(maybeepinfo)
+            self.states = states
+            self.dones = dones
+            self.obs = obs
             mb_rewards.append(rewards)
-
         mb_dones.append(self.dones)
 
         # Batch of steps to batch of rollouts
-        mb_obs = sf01(np.asarray(mb_obs, dtype=self.obs.dtype))
+        mb_obs = np.asarray(mb_obs, dtype=self.ob_dtype).swapaxes(1, 0).reshape(self.batch_ob_shape)
         mb_rewards = np.asarray(mb_rewards, dtype=np.float32).swapaxes(1, 0)
-        mb_actions = sf01(np.asarray(mb_actions, dtype=actions.dtype))
+        mb_actions = np.asarray(mb_actions, dtype=self.model.train_model.action.dtype.name).swapaxes(1, 0)
         mb_values = np.asarray(mb_values, dtype=np.float32).swapaxes(1, 0)
         mb_dones = np.asarray(mb_dones, dtype=np.bool).swapaxes(1, 0)
         mb_masks = mb_dones[:, :-1]
@@ -55,7 +57,7 @@ class Runner(AbstractEnvRunner):
 
         if self.gamma > 0.0:
             # Discount/bootstrap off value fn
-            last_values = self.model.value(tf.constant(self.obs))._numpy().tolist()
+            last_values = self.model.value(self.obs, S=self.states, M=self.dones).tolist()
             for n, (rewards, dones, value) in enumerate(zip(mb_rewards, mb_dones, last_values)):
                 rewards = rewards.tolist()
                 dones = dones.tolist()
@@ -66,15 +68,9 @@ class Runner(AbstractEnvRunner):
 
                 mb_rewards[n] = rewards
 
+        mb_actions = mb_actions.reshape(self.batch_action_shape)
 
         mb_rewards = mb_rewards.flatten()
         mb_values = mb_values.flatten()
         mb_masks = mb_masks.flatten()
         return mb_obs, mb_states, mb_rewards, mb_masks, mb_actions, mb_values, epinfos
-
-def sf01(arr):
-    """
-    swap and then flatten axes 0 and 1
-    """
-    s = arr.shape
-    return arr.swapaxes(0, 1).reshape(s[0] * s[1], *s[2:])

baselines/a2c/utils.py

@@ -1,5 +1,21 @@
+import os
 import numpy as np
 import tensorflow as tf
+from collections import deque
+
+def sample(logits):
+    noise = tf.random_uniform(tf.shape(logits))
+    return tf.argmax(logits - tf.log(-tf.log(noise)), 1)
+
+def cat_entropy(logits):
+    a0 = logits - tf.reduce_max(logits, 1, keepdims=True)
+    ea0 = tf.exp(a0)
+    z0 = tf.reduce_sum(ea0, 1, keepdims=True)
+    p0 = ea0 / z0
+    return tf.reduce_sum(p0 * (tf.log(z0) - a0), 1)
+
+def cat_entropy_softmax(p0):
+    return - tf.reduce_sum(p0 * tf.log(p0 + 1e-6), axis = 1)
 
 def ortho_init(scale=1.0):
     def _ortho_init(shape, dtype, partition_info=None):
@@ -18,18 +34,115 @@ def ortho_init(scale=1.0):
         return (scale * q[:shape[0], :shape[1]]).astype(np.float32)
     return _ortho_init
 
-def conv(scope, *, nf, rf, stride, activation, pad='valid', init_scale=1.0, data_format='channels_last'):
-    with tf.name_scope(scope):
-        layer = tf.keras.layers.Conv2D(filters=nf, kernel_size=rf, strides=stride, padding=pad,
-                                       data_format=data_format, kernel_initializer=ortho_init(init_scale))
-    return layer
+def conv(x, scope, *, nf, rf, stride, pad='VALID', init_scale=1.0, data_format='NHWC', one_dim_bias=False):
+    if data_format == 'NHWC':
+        channel_ax = 3
+        strides = [1, stride, stride, 1]
+        bshape = [1, 1, 1, nf]
+    elif data_format == 'NCHW':
+        channel_ax = 1
+        strides = [1, 1, stride, stride]
+        bshape = [1, nf, 1, 1]
+    else:
+        raise NotImplementedError
+    bias_var_shape = [nf] if one_dim_bias else [1, nf, 1, 1]
+    nin = x.get_shape()[channel_ax].value
+    wshape = [rf, rf, nin, nf]
+    with tf.variable_scope(scope):
+        w = tf.get_variable("w", wshape, initializer=ortho_init(init_scale))
+        b = tf.get_variable("b", bias_var_shape, initializer=tf.constant_initializer(0.0))
+        if not one_dim_bias and data_format == 'NHWC':
+            b = tf.reshape(b, bshape)
+        return tf.nn.conv2d(x, w, strides=strides, padding=pad, data_format=data_format) + b
 
-def fc(input_shape, scope, nh, *, init_scale=1.0, init_bias=0.0):
-    with tf.name_scope(scope):
-        layer = tf.keras.layers.Dense(units=nh, kernel_initializer=ortho_init(init_scale),
-                                      bias_initializer=tf.keras.initializers.Constant(init_bias))
-        layer.build(input_shape)
-    return layer
+def fc(x, scope, nh, *, init_scale=1.0, init_bias=0.0):
+    with tf.variable_scope(scope):
+        nin = x.get_shape()[1].value
+        w = tf.get_variable("w", [nin, nh], initializer=ortho_init(init_scale))
+        b = tf.get_variable("b", [nh], initializer=tf.constant_initializer(init_bias))
+        return tf.matmul(x, w)+b
+
+def batch_to_seq(h, nbatch, nsteps, flat=False):
+    if flat:
+        h = tf.reshape(h, [nbatch, nsteps])
+    else:
+        h = tf.reshape(h, [nbatch, nsteps, -1])
+    return [tf.squeeze(v, [1]) for v in tf.split(axis=1, num_or_size_splits=nsteps, value=h)]
+
+def seq_to_batch(h, flat = False):
+    shape = h[0].get_shape().as_list()
+    if not flat:
+        assert(len(shape) > 1)
+        nh = h[0].get_shape()[-1].value
+        return tf.reshape(tf.concat(axis=1, values=h), [-1, nh])
+    else:
+        return tf.reshape(tf.stack(values=h, axis=1), [-1])
+
+def lstm(xs, ms, s, scope, nh, init_scale=1.0):
+    nbatch, nin = [v.value for v in xs[0].get_shape()]
+    with tf.variable_scope(scope):
+        wx = tf.get_variable("wx", [nin, nh*4], initializer=ortho_init(init_scale))
+        wh = tf.get_variable("wh", [nh, nh*4], initializer=ortho_init(init_scale))
+        b = tf.get_variable("b", [nh*4], initializer=tf.constant_initializer(0.0))
+
+    c, h = tf.split(axis=1, num_or_size_splits=2, value=s)
+    for idx, (x, m) in enumerate(zip(xs, ms)):
+        c = c*(1-m)
+        h = h*(1-m)
+        z = tf.matmul(x, wx) + tf.matmul(h, wh) + b
+        i, f, o, u = tf.split(axis=1, num_or_size_splits=4, value=z)
+        i = tf.nn.sigmoid(i)
+        f = tf.nn.sigmoid(f)
+        o = tf.nn.sigmoid(o)
+        u = tf.tanh(u)
+        c = f*c + i*u
+        h = o*tf.tanh(c)
+        xs[idx] = h
+    s = tf.concat(axis=1, values=[c, h])
+    return xs, s
+
+def _ln(x, g, b, e=1e-5, axes=[1]):
+    u, s = tf.nn.moments(x, axes=axes, keep_dims=True)
+    x = (x-u)/tf.sqrt(s+e)
+    x = x*g+b
+    return x
+
+def lnlstm(xs, ms, s, scope, nh, init_scale=1.0):
+    nbatch, nin = [v.value for v in xs[0].get_shape()]
+    with tf.variable_scope(scope):
+        wx = tf.get_variable("wx", [nin, nh*4], initializer=ortho_init(init_scale))
+        gx = tf.get_variable("gx", [nh*4], initializer=tf.constant_initializer(1.0))
+        bx = tf.get_variable("bx", [nh*4], initializer=tf.constant_initializer(0.0))
+
+        wh = tf.get_variable("wh", [nh, nh*4], initializer=ortho_init(init_scale))
+        gh = tf.get_variable("gh", [nh*4], initializer=tf.constant_initializer(1.0))
+        bh = tf.get_variable("bh", [nh*4], initializer=tf.constant_initializer(0.0))
+
+        b = tf.get_variable("b", [nh*4], initializer=tf.constant_initializer(0.0))
+
+        gc = tf.get_variable("gc", [nh], initializer=tf.constant_initializer(1.0))
+        bc = tf.get_variable("bc", [nh], initializer=tf.constant_initializer(0.0))
+
+    c, h = tf.split(axis=1, num_or_size_splits=2, value=s)
+    for idx, (x, m) in enumerate(zip(xs, ms)):
+        c = c*(1-m)
+        h = h*(1-m)
+        z = _ln(tf.matmul(x, wx), gx, bx) + _ln(tf.matmul(h, wh), gh, bh) + b
+        i, f, o, u = tf.split(axis=1, num_or_size_splits=4, value=z)
+        i = tf.nn.sigmoid(i)
+        f = tf.nn.sigmoid(f)
+        o = tf.nn.sigmoid(o)
+        u = tf.tanh(u)
+        c = f*c + i*u
+        h = o*tf.tanh(_ln(c, gc, bc))
+        xs[idx] = h
+    s = tf.concat(axis=1, values=[c, h])
+    return xs, s
+
+def conv_to_fc(x):
+    nh = np.prod([v.value for v in x.get_shape()[1:]])
+    x = tf.reshape(x, [-1, nh])
+    return x
 
 def discount_with_dones(rewards, dones, gamma):
     discounted = []
@@ -39,25 +152,131 @@ def discount_with_dones(rewards, dones, gamma):
         discounted.append(r)
     return discounted[::-1]
 
-class InverseLinearTimeDecay(tf.keras.optimizers.schedules.LearningRateSchedule):
-    def __init__(self, initial_learning_rate, nupdates, name="InverseLinearTimeDecay"):
-        super(InverseLinearTimeDecay, self).__init__()
-        self.initial_learning_rate = initial_learning_rate
-        self.nupdates = nupdates
-        self.name = name
+def find_trainable_variables(key):
+    return tf.trainable_variables(key)
 
-    def __call__(self, step):
-        with tf.name_scope(self.name):
-            initial_learning_rate = tf.convert_to_tensor(self.initial_learning_rate, name="initial_learning_rate")
-            dtype = initial_learning_rate.dtype
-            step_t = tf.cast(step, dtype)
-            nupdates_t = tf.convert_to_tensor(self.nupdates, dtype=dtype)
-            tf.assert_less(step_t, nupdates_t)
-            return initial_learning_rate * (1. - step_t / nupdates_t)
+def make_path(f):
+    return os.makedirs(f, exist_ok=True)
 
-    def get_config(self):
-        return {
-            "initial_learning_rate": self.initial_learning_rate,
-            "nupdates": self.nupdates,
-            "name": self.name
-        }
+def constant(p):
+    return 1
+
+def linear(p):
+    return 1-p
+
+def middle_drop(p):
+    eps = 0.75
+    if 1-p<eps:
+        return eps*0.1
+    return 1-p
+
+def double_linear_con(p):
+    p *= 2
+    eps = 0.125
+    if 1-p<eps:
+        return eps
+    return 1-p
+
+def double_middle_drop(p):
+    eps1 = 0.75
+    eps2 = 0.25
+    if 1-p<eps1:
+        if 1-p<eps2:
+            return eps2*0.5
+        return eps1*0.1
+    return 1-p
+
+schedules = {
+    'linear':linear,
+    'constant':constant,
+    'double_linear_con': double_linear_con,
+    'middle_drop': middle_drop,
+    'double_middle_drop': double_middle_drop
+}
+
+class Scheduler(object):
+
+    def __init__(self, v, nvalues, schedule):
+        self.n = 0.
+        self.v = v
+        self.nvalues = nvalues
+        self.schedule = schedules[schedule]
+
+    def value(self):
+        current_value = self.v*self.schedule(self.n/self.nvalues)
+        self.n += 1.
+        return current_value
+
+    def value_steps(self, steps):
+        return self.v*self.schedule(steps/self.nvalues)
+
+
+class EpisodeStats:
+    def __init__(self, nsteps, nenvs):
+        self.episode_rewards = []
+        for i in range(nenvs):
+            self.episode_rewards.append([])
+        self.lenbuffer = deque(maxlen=40)  # rolling buffer for episode lengths
+        self.rewbuffer = deque(maxlen=40)  # rolling buffer for episode rewards
+        self.nsteps = nsteps
+        self.nenvs = nenvs
+
+    def feed(self, rewards, masks):
+        rewards = np.reshape(rewards, [self.nenvs, self.nsteps])
+        masks = np.reshape(masks, [self.nenvs, self.nsteps])
+        for i in range(0, self.nenvs):
+            for j in range(0, self.nsteps):
+                self.episode_rewards[i].append(rewards[i][j])
+                if masks[i][j]:
+                    l = len(self.episode_rewards[i])
+                    s = sum(self.episode_rewards[i])
+                    self.lenbuffer.append(l)
+                    self.rewbuffer.append(s)
+                    self.episode_rewards[i] = []
+
+    def mean_length(self):
+        if self.lenbuffer:
+            return np.mean(self.lenbuffer)
+        else:
+            return 0  # on the first params dump, no episodes are finished
+
+    def mean_reward(self):
+        if self.rewbuffer:
+            return np.mean(self.rewbuffer)
+        else:
+            return 0
+
+
+# For ACER
+def get_by_index(x, idx):
+    assert(len(x.get_shape()) == 2)
+    assert(len(idx.get_shape()) == 1)
+    idx_flattened = tf.range(0, x.shape[0]) * x.shape[1] + idx
+    y = tf.gather(tf.reshape(x, [-1]),  # flatten input
+                  idx_flattened)  # use flattened indices
+    return y
+
+def check_shape(ts,shapes):
+    i = 0
+    for (t,shape) in zip(ts,shapes):
+        assert t.get_shape().as_list()==shape, "id " + str(i) + " shape " + str(t.get_shape()) + str(shape)
+        i += 1
+
+def avg_norm(t):
+    return tf.reduce_mean(tf.sqrt(tf.reduce_sum(tf.square(t), axis=-1)))
+
+def gradient_add(g1, g2, param):
+    print([g1, g2, param.name])
+    assert (not (g1 is None and g2 is None)), param.name
+    if g1 is None:
+        return g2
+    elif g2 is None:
+        return g1
+    else:
+        return g1 + g2
+
+def q_explained_variance(qpred, q):
+    _, vary = tf.nn.moments(q, axes=[0, 1])
+    _, varpred = tf.nn.moments(q - qpred, axes=[0, 1])
+    check_shape([vary, varpred], [[]] * 2)
+    return 1.0 - (varpred / vary)

baselines/acer/README.md

@@ -0,0 +1,6 @@
+# ACER
+
+- Original paper: https://arxiv.org/abs/1611.01224
+- `python -m baselines.run --alg=acer --env=PongNoFrameskip-v4` runs the algorithm for 40M frames = 10M timesteps on an Atari Pong. See help (`-h`) for more options.
+- also refer to the repo-wide [README.md](../../README.md#training-models)
+

baselines/acer/acer.py

@@ -0,0 +1,377 @@
+import time
+import functools
+import numpy as np
+import tensorflow as tf
+from baselines import logger
+
+from baselines.common import set_global_seeds
+from baselines.common.policies import build_policy
+from baselines.common.tf_util import get_session, save_variables
+from baselines.common.vec_env.vec_frame_stack import VecFrameStack
+
+from baselines.a2c.utils import batch_to_seq, seq_to_batch
+from baselines.a2c.utils import cat_entropy_softmax
+from baselines.a2c.utils import Scheduler, find_trainable_variables
+from baselines.a2c.utils import EpisodeStats
+from baselines.a2c.utils import get_by_index, check_shape, avg_norm, gradient_add, q_explained_variance
+from baselines.acer.buffer import Buffer
+from baselines.acer.runner import Runner
+
+# remove last step
+def strip(var, nenvs, nsteps, flat = False):
+    vars = batch_to_seq(var, nenvs, nsteps + 1, flat)
+    return seq_to_batch(vars[:-1], flat)
+
+def q_retrace(R, D, q_i, v, rho_i, nenvs, nsteps, gamma):
+    """
+    Calculates q_retrace targets
+
+    :param R: Rewards
+    :param D: Dones
+    :param q_i: Q values for actions taken
+    :param v: V values
+    :param rho_i: Importance weight for each action
+    :return: Q_retrace values
+    """
+    rho_bar = batch_to_seq(tf.minimum(1.0, rho_i), nenvs, nsteps, True)  # list of len steps, shape [nenvs]
+    rs = batch_to_seq(R, nenvs, nsteps, True)  # list of len steps, shape [nenvs]
+    ds = batch_to_seq(D, nenvs, nsteps, True)  # list of len steps, shape [nenvs]
+    q_is = batch_to_seq(q_i, nenvs, nsteps, True)
+    vs = batch_to_seq(v, nenvs, nsteps + 1, True)
+    v_final = vs[-1]
+    qret = v_final
+    qrets = []
+    for i in range(nsteps - 1, -1, -1):
+        check_shape([qret, ds[i], rs[i], rho_bar[i], q_is[i], vs[i]], [[nenvs]] * 6)
+        qret = rs[i] + gamma * qret * (1.0 - ds[i])
+        qrets.append(qret)
+        qret = (rho_bar[i] * (qret - q_is[i])) + vs[i]
+    qrets = qrets[::-1]
+    qret = seq_to_batch(qrets, flat=True)
+    return qret
+
+# For ACER with PPO clipping instead of trust region
+# def clip(ratio, eps_clip):
+#     # assume 0 <= eps_clip <= 1
+#     return tf.minimum(1 + eps_clip, tf.maximum(1 - eps_clip, ratio))
+
+class Model(object):
+    def __init__(self, policy, ob_space, ac_space, nenvs, nsteps, ent_coef, q_coef, gamma, max_grad_norm, lr,
+                 rprop_alpha, rprop_epsilon, total_timesteps, lrschedule,
+                 c, trust_region, alpha, delta):
+
+        sess = get_session()
+        nact = ac_space.n
+        nbatch = nenvs * nsteps
+
+        A = tf.placeholder(tf.int32, [nbatch]) # actions
+        D = tf.placeholder(tf.float32, [nbatch]) # dones
+        R = tf.placeholder(tf.float32, [nbatch]) # rewards, not returns
+        MU = tf.placeholder(tf.float32, [nbatch, nact]) # mu's
+        LR = tf.placeholder(tf.float32, [])
+        eps = 1e-6
+
+        step_ob_placeholder = tf.placeholder(dtype=ob_space.dtype, shape=(nenvs,) + ob_space.shape)
+        train_ob_placeholder = tf.placeholder(dtype=ob_space.dtype, shape=(nenvs*(nsteps+1),) + ob_space.shape)
+        with tf.variable_scope('acer_model', reuse=tf.AUTO_REUSE):
+
+            step_model = policy(nbatch=nenvs, nsteps=1, observ_placeholder=step_ob_placeholder, sess=sess)
+            train_model = policy(nbatch=nbatch, nsteps=nsteps, observ_placeholder=train_ob_placeholder, sess=sess)
+
+
+        params = find_trainable_variables("acer_model")
+        print("Params {}".format(len(params)))
+        for var in params:
+            print(var)
+
+        # create polyak averaged model
+        ema = tf.train.ExponentialMovingAverage(alpha)
+        ema_apply_op = ema.apply(params)
+
+        def custom_getter(getter, *args, **kwargs):
+            v = ema.average(getter(*args, **kwargs))
+            print(v.name)
+            return v
+
+        with tf.variable_scope("acer_model", custom_getter=custom_getter, reuse=True):
+            polyak_model = policy(nbatch=nbatch, nsteps=nsteps, observ_placeholder=train_ob_placeholder, sess=sess)
+
+        # Notation: (var) = batch variable, (var)s = seqeuence variable, (var)_i = variable index by action at step i
+
+        # action probability distributions according to train_model, polyak_model and step_model
+        # poilcy.pi is probability distribution parameters; to obtain distribution that sums to 1 need to take softmax
+        train_model_p = tf.nn.softmax(train_model.pi)
+        polyak_model_p = tf.nn.softmax(polyak_model.pi)
+        step_model_p = tf.nn.softmax(step_model.pi)
+        v = tf.reduce_sum(train_model_p * train_model.q, axis = -1) # shape is [nenvs * (nsteps + 1)]
+
+        # strip off last step
+        f, f_pol, q = map(lambda var: strip(var, nenvs, nsteps), [train_model_p, polyak_model_p, train_model.q])
+        # Get pi and q values for actions taken
+        f_i = get_by_index(f, A)
+        q_i = get_by_index(q, A)
+
+        # Compute ratios for importance truncation
+        rho = f / (MU + eps)
+        rho_i = get_by_index(rho, A)
+
+        # Calculate Q_retrace targets
+        qret = q_retrace(R, D, q_i, v, rho_i, nenvs, nsteps, gamma)
+
+        # Calculate losses
+        # Entropy
+        # entropy = tf.reduce_mean(strip(train_model.pd.entropy(), nenvs, nsteps))
+        entropy = tf.reduce_mean(cat_entropy_softmax(f))
+
+        # Policy Graident loss, with truncated importance sampling & bias correction
+        v = strip(v, nenvs, nsteps, True)
+        check_shape([qret, v, rho_i, f_i], [[nenvs * nsteps]] * 4)
+        check_shape([rho, f, q], [[nenvs * nsteps, nact]] * 2)
+
+        # Truncated importance sampling
+        adv = qret - v
+        logf = tf.log(f_i + eps)
+        gain_f = logf * tf.stop_gradient(adv * tf.minimum(c, rho_i))  # [nenvs * nsteps]
+        loss_f = -tf.reduce_mean(gain_f)
+
+        # Bias correction for the truncation
+        adv_bc = (q - tf.reshape(v, [nenvs * nsteps, 1]))  # [nenvs * nsteps, nact]
+        logf_bc = tf.log(f + eps) # / (f_old + eps)
+        check_shape([adv_bc, logf_bc], [[nenvs * nsteps, nact]]*2)
+        gain_bc = tf.reduce_sum(logf_bc * tf.stop_gradient(adv_bc * tf.nn.relu(1.0 - (c / (rho + eps))) * f), axis = 1) #IMP: This is sum, as expectation wrt f
+        loss_bc= -tf.reduce_mean(gain_bc)
+
+        loss_policy = loss_f + loss_bc
+
+        # Value/Q function loss, and explained variance
+        check_shape([qret, q_i], [[nenvs * nsteps]]*2)
+        ev = q_explained_variance(tf.reshape(q_i, [nenvs, nsteps]), tf.reshape(qret, [nenvs, nsteps]))
+        loss_q = tf.reduce_mean(tf.square(tf.stop_gradient(qret) - q_i)*0.5)
+
+        # Net loss
+        check_shape([loss_policy, loss_q, entropy], [[]] * 3)
+        loss = loss_policy + q_coef * loss_q - ent_coef * entropy
+
+        if trust_region:
+            g = tf.gradients(- (loss_policy - ent_coef * entropy) * nsteps * nenvs, f) #[nenvs * nsteps, nact]
+            # k = tf.gradients(KL(f_pol || f), f)
+            k = - f_pol / (f + eps) #[nenvs * nsteps, nact] # Directly computed gradient of KL divergence wrt f
+            k_dot_g = tf.reduce_sum(k * g, axis=-1)
+            adj = tf.maximum(0.0, (tf.reduce_sum(k * g, axis=-1) - delta) / (tf.reduce_sum(tf.square(k), axis=-1) + eps)) #[nenvs * nsteps]
+
+            # Calculate stats (before doing adjustment) for logging.
+            avg_norm_k = avg_norm(k)
+            avg_norm_g = avg_norm(g)
+            avg_norm_k_dot_g = tf.reduce_mean(tf.abs(k_dot_g))
+            avg_norm_adj = tf.reduce_mean(tf.abs(adj))
+
+            g = g - tf.reshape(adj, [nenvs * nsteps, 1]) * k
+            grads_f = -g/(nenvs*nsteps) # These are turst region adjusted gradients wrt f ie statistics of policy pi
+            grads_policy = tf.gradients(f, params, grads_f)
+            grads_q = tf.gradients(loss_q * q_coef, params)
+            grads = [gradient_add(g1, g2, param) for (g1, g2, param) in zip(grads_policy, grads_q, params)]
+
+            avg_norm_grads_f = avg_norm(grads_f) * (nsteps * nenvs)
+            norm_grads_q = tf.global_norm(grads_q)
+            norm_grads_policy = tf.global_norm(grads_policy)
+        else:
+            grads = tf.gradients(loss, params)
+
+        if max_grad_norm is not None:
+            grads, norm_grads = tf.clip_by_global_norm(grads, max_grad_norm)
+        grads = list(zip(grads, params))
+        trainer = tf.train.RMSPropOptimizer(learning_rate=LR, decay=rprop_alpha, epsilon=rprop_epsilon)
+        _opt_op = trainer.apply_gradients(grads)
+
+        # so when you call _train, you first do the gradient step, then you apply ema
+        with tf.control_dependencies([_opt_op]):
+            _train = tf.group(ema_apply_op)
+
+        lr = Scheduler(v=lr, nvalues=total_timesteps, schedule=lrschedule)
+
+        # Ops/Summaries to run, and their names for logging
+        run_ops = [_train, loss, loss_q, entropy, loss_policy, loss_f, loss_bc, ev, norm_grads]
+        names_ops = ['loss', 'loss_q', 'entropy', 'loss_policy', 'loss_f', 'loss_bc', 'explained_variance',
+                     'norm_grads']
+        if trust_region:
+            run_ops = run_ops + [norm_grads_q, norm_grads_policy, avg_norm_grads_f, avg_norm_k, avg_norm_g, avg_norm_k_dot_g,
+                                 avg_norm_adj]
+            names_ops = names_ops + ['norm_grads_q', 'norm_grads_policy', 'avg_norm_grads_f', 'avg_norm_k', 'avg_norm_g',
+                                     'avg_norm_k_dot_g', 'avg_norm_adj']
+
+        def train(obs, actions, rewards, dones, mus, states, masks, steps):
+            cur_lr = lr.value_steps(steps)
+            td_map = {train_model.X: obs, polyak_model.X: obs, A: actions, R: rewards, D: dones, MU: mus, LR: cur_lr}
+            if states is not None:
+                td_map[train_model.S] = states
+                td_map[train_model.M] = masks
+                td_map[polyak_model.S] = states
+                td_map[polyak_model.M] = masks
+
+            return names_ops, sess.run(run_ops, td_map)[1:]  # strip off _train
+
+        def _step(observation, **kwargs):
+            return step_model._evaluate([step_model.action, step_model_p, step_model.state], observation, **kwargs)
+
+
+
+        self.train = train
+        self.save = functools.partial(save_variables, sess=sess, variables=params)
+        self.train_model = train_model
+        self.step_model = step_model
+        self._step = _step
+        self.step = self.step_model.step
+
+        self.initial_state = step_model.initial_state
+        tf.global_variables_initializer().run(session=sess)
+
+
+class Acer():
+    def __init__(self, runner, model, buffer, log_interval):
+        self.runner = runner
+        self.model = model
+        self.buffer = buffer
+        self.log_interval = log_interval
+        self.tstart = None
+        self.episode_stats = EpisodeStats(runner.nsteps, runner.nenv)
+        self.steps = None
+
+    def call(self, on_policy):
+        runner, model, buffer, steps = self.runner, self.model, self.buffer, self.steps
+        if on_policy:
+            enc_obs, obs, actions, rewards, mus, dones, masks = runner.run()
+            self.episode_stats.feed(rewards, dones)
+            if buffer is not None:
+                buffer.put(enc_obs, actions, rewards, mus, dones, masks)
+        else:
+            # get obs, actions, rewards, mus, dones from buffer.
+            obs, actions, rewards, mus, dones, masks = buffer.get()
+
+
+        # reshape stuff correctly
+        obs = obs.reshape(runner.batch_ob_shape)
+        actions = actions.reshape([runner.nbatch])
+        rewards = rewards.reshape([runner.nbatch])
+        mus = mus.reshape([runner.nbatch, runner.nact])
+        dones = dones.reshape([runner.nbatch])
+        masks = masks.reshape([runner.batch_ob_shape[0]])
+
+        names_ops, values_ops = model.train(obs, actions, rewards, dones, mus, model.initial_state, masks, steps)
+
+        if on_policy and (int(steps/runner.nbatch) % self.log_interval == 0):
+            logger.record_tabular("total_timesteps", steps)
+            logger.record_tabular("fps", int(steps/(time.time() - self.tstart)))
+            # IMP: In EpisodicLife env, during training, we get done=True at each loss of life, not just at the terminal state.
+            # Thus, this is mean until end of life, not end of episode.
+            # For true episode rewards, see the monitor files in the log folder.
+            logger.record_tabular("mean_episode_length", self.episode_stats.mean_length())
+            logger.record_tabular("mean_episode_reward", self.episode_stats.mean_reward())
+            for name, val in zip(names_ops, values_ops):
+                logger.record_tabular(name, float(val))
+            logger.dump_tabular()
+
+
+def learn(network, env, seed=None, nsteps=20, total_timesteps=int(80e6), q_coef=0.5, ent_coef=0.01,
+          max_grad_norm=10, lr=7e-4, lrschedule='linear', rprop_epsilon=1e-5, rprop_alpha=0.99, gamma=0.99,
+          log_interval=100, buffer_size=50000, replay_ratio=4, replay_start=10000, c=10.0,
+          trust_region=True, alpha=0.99, delta=1, load_path=None, **network_kwargs):
+
+    '''
+    Main entrypoint for ACER (Actor-Critic with Experience Replay) algorithm (https://arxiv.org/pdf/1611.01224.pdf)
+    Train an agent with given network architecture on a given environment using ACER.
+
+    Parameters:
+    ----------
+
+    network:            policy network architecture. Either string (mlp, lstm, lnlstm, cnn_lstm, cnn, cnn_small, conv_only - see baselines.common/models.py for full list)
+                        specifying the standard network architecture, or a function that takes tensorflow tensor as input and returns
+                        tuple (output_tensor, extra_feed) where output tensor is the last network layer output, extra_feed is None for feed-forward
+                        neural nets, and extra_feed is a dictionary describing how to feed state into the network for recurrent neural nets.
+                        See baselines.common/policies.py/lstm for more details on using recurrent nets in policies
+
+    env:                environment. Needs to be vectorized for parallel environment simulation.
+                        The environments produced by gym.make can be wrapped using baselines.common.vec_env.DummyVecEnv class.
+
+    nsteps:             int, number of steps of the vectorized environment per update (i.e. batch size is nsteps * nenv where
+                        nenv is number of environment copies simulated in parallel) (default: 20)
+
+    nstack:             int, size of the frame stack, i.e. number of the frames passed to the step model. Frames are stacked along channel dimension
+                        (last image dimension) (default: 4)
+
+    total_timesteps:    int, number of timesteps (i.e. number of actions taken in the environment) (default: 80M)
+
+    q_coef:             float, value function loss coefficient in the optimization objective (analog of vf_coef for other actor-critic methods)
+
+    ent_coef:           float, policy entropy coefficient in the optimization objective (default: 0.01)
+
+    max_grad_norm:      float, gradient norm clipping coefficient. If set to None, no clipping. (default: 10),
+
+    lr:                 float, learning rate for RMSProp (current implementation has RMSProp hardcoded in) (default: 7e-4)
+
+    lrschedule:         schedule of learning rate. Can be 'linear', 'constant', or a function [0..1] -> [0..1] that takes fraction of the training progress as input and
+                        returns fraction of the learning rate (specified as lr) as output
+
+    rprop_epsilon:      float, RMSProp epsilon (stabilizes square root computation in denominator of RMSProp update) (default: 1e-5)
+
+    rprop_alpha:        float, RMSProp decay parameter (default: 0.99)
+
+    gamma:              float, reward discounting factor (default: 0.99)
+
+    log_interval:       int, number of updates between logging events (default: 100)
+
+    buffer_size:        int, size of the replay buffer (default: 50k)
+
+    replay_ratio:       int, now many (on average) batches of data to sample from the replay buffer take after batch from the environment (default: 4)
+
+    replay_start:       int, the sampling from the replay buffer does not start until replay buffer has at least that many samples (default: 10k)
+
+    c:                  float, importance weight clipping factor (default: 10)
+
+    trust_region        bool, whether or not algorithms estimates the gradient KL divergence between the old and updated policy and uses it to determine step size  (default: True)
+
+    delta:              float, max KL divergence between the old policy and updated policy (default: 1)
+
+    alpha:              float, momentum factor in the Polyak (exponential moving average) averaging of the model parameters (default: 0.99)
+
+    load_path:          str, path to load the model from (default: None)
+
+    **network_kwargs:               keyword arguments to the policy / network builder. See baselines.common/policies.py/build_policy and arguments to a particular type of network
+                                    For instance, 'mlp' network architecture has arguments num_hidden and num_layers.
+
+    '''
+
+    print("Running Acer Simple")
+    print(locals())
+    set_global_seeds(seed)
+    if not isinstance(env, VecFrameStack):
+        env = VecFrameStack(env, 1)
+
+    policy = build_policy(env, network, estimate_q=True, **network_kwargs)
+    nenvs = env.num_envs
+    ob_space = env.observation_space
+    ac_space = env.action_space
+
+    nstack = env.nstack
+    model = Model(policy=policy, ob_space=ob_space, ac_space=ac_space, nenvs=nenvs, nsteps=nsteps,
+                  ent_coef=ent_coef, q_coef=q_coef, gamma=gamma,
+                  max_grad_norm=max_grad_norm, lr=lr, rprop_alpha=rprop_alpha, rprop_epsilon=rprop_epsilon,
+                  total_timesteps=total_timesteps, lrschedule=lrschedule, c=c,
+                  trust_region=trust_region, alpha=alpha, delta=delta)
+
+    runner = Runner(env=env, model=model, nsteps=nsteps)
+    if replay_ratio > 0:
+        buffer = Buffer(env=env, nsteps=nsteps, size=buffer_size)
+    else:
+        buffer = None
+    nbatch = nenvs*nsteps
+    acer = Acer(runner, model, buffer, log_interval)
+    acer.tstart = time.time()
+
+    for acer.steps in range(0, total_timesteps, nbatch): #nbatch samples, 1 on_policy call and multiple off-policy calls
+        acer.call(on_policy=True)
+        if replay_ratio > 0 and buffer.has_atleast(replay_start):
+            n = np.random.poisson(replay_ratio)
+            for _ in range(n):
+                acer.call(on_policy=False)  # no simulation steps in this
+
+    return model

baselines/acer/buffer.py

@@ -0,0 +1,156 @@
+import numpy as np
+
+class Buffer(object):
+    # gets obs, actions, rewards, mu's, (states, masks), dones
+    def __init__(self, env, nsteps, size=50000):
+        self.nenv = env.num_envs
+        self.nsteps = nsteps
+        # self.nh, self.nw, self.nc = env.observation_space.shape
+        self.obs_shape = env.observation_space.shape
+        self.obs_dtype = env.observation_space.dtype
+        self.ac_dtype = env.action_space.dtype
+        self.nc = self.obs_shape[-1]
+        self.nstack = env.nstack
+        self.nc //= self.nstack
+        self.nbatch = self.nenv * self.nsteps
+        self.size = size // (self.nsteps)  # Each loc contains nenv * nsteps frames, thus total buffer is nenv * size frames
+
+        # Memory
+        self.enc_obs = None
+        self.actions = None
+        self.rewards = None
+        self.mus = None
+        self.dones = None
+        self.masks = None
+
+        # Size indexes
+        self.next_idx = 0
+        self.num_in_buffer = 0
+
+    def has_atleast(self, frames):
+        # Frames per env, so total (nenv * frames) Frames needed
+        # Each buffer loc has nenv * nsteps frames
+        return self.num_in_buffer >= (frames // self.nsteps)
+
+    def can_sample(self):
+        return self.num_in_buffer > 0
+
+    # Generate stacked frames
+    def decode(self, enc_obs, dones):
+        # enc_obs has shape [nenvs, nsteps + nstack, nh, nw, nc]
+        # dones has shape [nenvs, nsteps]
+        # returns stacked obs of shape [nenv, (nsteps + 1), nh, nw, nstack*nc]
+
+        return _stack_obs(enc_obs, dones,
+                          nsteps=self.nsteps)
+
+    def put(self, enc_obs, actions, rewards, mus, dones, masks):
+        # enc_obs [nenv, (nsteps + nstack), nh, nw, nc]
+        # actions, rewards, dones [nenv, nsteps]
+        # mus [nenv, nsteps, nact]
+
+        if self.enc_obs is None:
+            self.enc_obs = np.empty([self.size] + list(enc_obs.shape), dtype=self.obs_dtype)
+            self.actions = np.empty([self.size] + list(actions.shape), dtype=self.ac_dtype)
+            self.rewards = np.empty([self.size] + list(rewards.shape), dtype=np.float32)
+            self.mus = np.empty([self.size] + list(mus.shape), dtype=np.float32)
+            self.dones = np.empty([self.size] + list(dones.shape), dtype=np.bool)
+            self.masks = np.empty([self.size] + list(masks.shape), dtype=np.bool)
+
+        self.enc_obs[self.next_idx] = enc_obs
+        self.actions[self.next_idx] = actions
+        self.rewards[self.next_idx] = rewards
+        self.mus[self.next_idx] = mus
+        self.dones[self.next_idx] = dones
+        self.masks[self.next_idx] = masks
+
+        self.next_idx = (self.next_idx + 1) % self.size
+        self.num_in_buffer = min(self.size, self.num_in_buffer + 1)
+
+    def take(self, x, idx, envx):
+        nenv = self.nenv
+        out = np.empty([nenv] + list(x.shape[2:]), dtype=x.dtype)
+        for i in range(nenv):
+            out[i] = x[idx[i], envx[i]]
+        return out
+
+    def get(self):
+        # returns
+        # obs [nenv, (nsteps + 1), nh, nw, nstack*nc]
+        # actions, rewards, dones [nenv, nsteps]
+        # mus [nenv, nsteps, nact]
+        nenv = self.nenv
+        assert self.can_sample()
+
+        # Sample exactly one id per env. If you sample across envs, then higher correlation in samples from same env.
+        idx = np.random.randint(0, self.num_in_buffer, nenv)
+        envx = np.arange(nenv)
+
+        take = lambda x: self.take(x, idx, envx)  # for i in range(nenv)], axis = 0)
+        dones = take(self.dones)
+        enc_obs = take(self.enc_obs)
+        obs = self.decode(enc_obs, dones)
+        actions = take(self.actions)
+        rewards = take(self.rewards)
+        mus = take(self.mus)
+        masks = take(self.masks)
+        return obs, actions, rewards, mus, dones, masks
+
+
+
+def _stack_obs_ref(enc_obs, dones, nsteps):
+    nenv = enc_obs.shape[0]
+    nstack = enc_obs.shape[1] - nsteps
+    nh, nw, nc = enc_obs.shape[2:]
+    obs_dtype = enc_obs.dtype
+    obs_shape = (nh, nw, nc*nstack)
+
+    mask = np.empty([nsteps + nstack - 1, nenv, 1, 1, 1], dtype=np.float32)
+    obs = np.zeros([nstack, nsteps + nstack, nenv, nh, nw, nc], dtype=obs_dtype)
+    x = np.reshape(enc_obs, [nenv, nsteps + nstack, nh, nw, nc]).swapaxes(1, 0)  # [nsteps + nstack, nenv, nh, nw, nc]
+
+    mask[nstack-1:] = np.reshape(1.0 - dones, [nenv, nsteps, 1, 1, 1]).swapaxes(1, 0)  # keep
+    mask[:nstack-1] = 1.0
+
+    # y = np.reshape(1 - dones, [nenvs, nsteps, 1, 1, 1])
+    for i in range(nstack):
+        obs[-(i + 1), i:] = x
+        # obs[:,i:,:,:,-(i+1),:] = x
+        x = x[:-1] * mask
+        mask = mask[1:]
+
+    return np.reshape(obs[:, (nstack-1):].transpose((2, 1, 3, 4, 0, 5)), (nenv, (nsteps + 1)) + obs_shape)
+
+def _stack_obs(enc_obs, dones, nsteps):
+    nenv = enc_obs.shape[0]
+    nstack = enc_obs.shape[1] - nsteps
+    nc = enc_obs.shape[-1]
+
+    obs_ = np.zeros((nenv, nsteps + 1) + enc_obs.shape[2:-1] + (enc_obs.shape[-1] * nstack, ), dtype=enc_obs.dtype)
+    mask = np.ones((nenv, nsteps+1), dtype=enc_obs.dtype)
+    mask[:, 1:] = 1.0 - dones
+    mask = mask.reshape(mask.shape + tuple(np.ones(len(enc_obs.shape)-2, dtype=np.uint8)))
+
+    for i in range(nstack-1, -1, -1):
+        obs_[..., i * nc : (i + 1) * nc] = enc_obs[:, i : i + nsteps + 1, :]
+        if i < nstack-1:
+            obs_[..., i * nc : (i + 1) * nc] *= mask
+            mask[:, 1:, ...] *= mask[:, :-1, ...]
+
+    return obs_
+
+def test_stack_obs():
+    nstack = 7
+    nenv = 1
+    nsteps = 5
+
+    obs_shape = (2, 3, nstack)
+
+    enc_obs_shape = (nenv, nsteps + nstack) + obs_shape[:-1] + (1,)
+    enc_obs = np.random.random(enc_obs_shape)
+    dones = np.random.randint(low=0, high=2, size=(nenv, nsteps))
+
+    stacked_obs_ref = _stack_obs_ref(enc_obs, dones, nsteps=nsteps)
+    stacked_obs_test = _stack_obs(enc_obs, dones, nsteps=nsteps)
+
+    np.testing.assert_allclose(stacked_obs_ref, stacked_obs_test)

baselines/acer/defaults.py

@@ -0,0 +1,4 @@
+def atari():
+    return dict(
+        lrschedule='constant'
+    )

baselines/acer/policies.py

@@ -0,0 +1,81 @@
+import numpy as np
+import tensorflow as tf
+from baselines.common.policies import nature_cnn
+from baselines.a2c.utils import fc, batch_to_seq, seq_to_batch, lstm, sample
+
+
+class AcerCnnPolicy(object):
+
+    def __init__(self, sess, ob_space, ac_space, nenv, nsteps, nstack, reuse=False):
+        nbatch = nenv * nsteps
+        nh, nw, nc = ob_space.shape
+        ob_shape = (nbatch, nh, nw, nc * nstack)
+        nact = ac_space.n
+        X = tf.placeholder(tf.uint8, ob_shape)  # obs
+        with tf.variable_scope("model", reuse=reuse):
+            h = nature_cnn(X)
+            pi_logits = fc(h, 'pi', nact, init_scale=0.01)
+            pi = tf.nn.softmax(pi_logits)
+            q = fc(h, 'q', nact)
+
+        a = sample(tf.nn.softmax(pi_logits))  # could change this to use self.pi instead
+        self.initial_state = []  # not stateful
+        self.X = X
+        self.pi = pi  # actual policy params now
+        self.pi_logits = pi_logits
+        self.q = q
+        self.vf = q
+
+        def step(ob, *args, **kwargs):
+            # returns actions, mus, states
+            a0, pi0 = sess.run([a, pi], {X: ob})
+            return a0, pi0, []  # dummy state
+
+        def out(ob, *args, **kwargs):
+            pi0, q0 = sess.run([pi, q], {X: ob})
+            return pi0, q0
+
+        def act(ob, *args, **kwargs):
+            return sess.run(a, {X: ob})
+
+        self.step = step
+        self.out = out
+        self.act = act
+
+class AcerLstmPolicy(object):
+
+    def __init__(self, sess, ob_space, ac_space, nenv, nsteps, nstack, reuse=False, nlstm=256):
+        nbatch = nenv * nsteps
+        nh, nw, nc = ob_space.shape
+        ob_shape = (nbatch, nh, nw, nc * nstack)
+        nact = ac_space.n
+        X = tf.placeholder(tf.uint8, ob_shape)  # obs
+        M = tf.placeholder(tf.float32, [nbatch]) #mask (done t-1)
+        S = tf.placeholder(tf.float32, [nenv, nlstm*2]) #states
+        with tf.variable_scope("model", reuse=reuse):
+            h = nature_cnn(X)
+
+            # lstm
+            xs = batch_to_seq(h, nenv, nsteps)
+            ms = batch_to_seq(M, nenv, nsteps)
+            h5, snew = lstm(xs, ms, S, 'lstm1', nh=nlstm)
+            h5 = seq_to_batch(h5)
+
+            pi_logits = fc(h5, 'pi', nact, init_scale=0.01)
+            pi = tf.nn.softmax(pi_logits)
+            q = fc(h5, 'q', nact)
+
+        a = sample(pi_logits)  # could change this to use self.pi instead
+        self.initial_state = np.zeros((nenv, nlstm*2), dtype=np.float32)
+        self.X = X
+        self.M = M
+        self.S = S
+        self.pi = pi  # actual policy params now
+        self.q = q
+
+        def step(ob, state, mask, *args, **kwargs):
+            # returns actions, mus, states
+            a0, pi0, s = sess.run([a, pi, snew], {X: ob, S: state, M: mask})
+            return a0, pi0, s
+
+        self.step = step

baselines/acer/runner.py

@@ -0,0 +1,61 @@
+import numpy as np
+from baselines.common.runners import AbstractEnvRunner
+from baselines.common.vec_env.vec_frame_stack import VecFrameStack
+from gym import spaces
+
+
+class Runner(AbstractEnvRunner):
+
+    def __init__(self, env, model, nsteps):
+        super().__init__(env=env, model=model, nsteps=nsteps)
+        assert isinstance(env.action_space, spaces.Discrete), 'This ACER implementation works only with discrete action spaces!'
+        assert isinstance(env, VecFrameStack)
+
+        self.nact = env.action_space.n
+        nenv = self.nenv
+        self.nbatch = nenv * nsteps
+        self.batch_ob_shape = (nenv*(nsteps+1),) + env.observation_space.shape
+
+        self.obs = env.reset()
+        self.obs_dtype = env.observation_space.dtype
+        self.ac_dtype = env.action_space.dtype
+        self.nstack = self.env.nstack
+        self.nc = self.batch_ob_shape[-1] // self.nstack
+
+
+    def run(self):
+        # enc_obs = np.split(self.obs, self.nstack, axis=3)  # so now list of obs steps
+        enc_obs = np.split(self.env.stackedobs, self.env.nstack, axis=-1)
+        mb_obs, mb_actions, mb_mus, mb_dones, mb_rewards = [], [], [], [], []
+        for _ in range(self.nsteps):
+            actions, mus, states = self.model._step(self.obs, S=self.states, M=self.dones)
+            mb_obs.append(np.copy(self.obs))
+            mb_actions.append(actions)
+            mb_mus.append(mus)
+            mb_dones.append(self.dones)
+            obs, rewards, dones, _ = self.env.step(actions)
+            # states information for statefull models like LSTM
+            self.states = states
+            self.dones = dones
+            self.obs = obs
+            mb_rewards.append(rewards)
+            enc_obs.append(obs[..., -self.nc:])
+        mb_obs.append(np.copy(self.obs))
+        mb_dones.append(self.dones)
+
+        enc_obs = np.asarray(enc_obs, dtype=self.obs_dtype).swapaxes(1, 0)
+        mb_obs = np.asarray(mb_obs, dtype=self.obs_dtype).swapaxes(1, 0)
+        mb_actions = np.asarray(mb_actions, dtype=self.ac_dtype).swapaxes(1, 0)
+        mb_rewards = np.asarray(mb_rewards, dtype=np.float32).swapaxes(1, 0)
+        mb_mus = np.asarray(mb_mus, dtype=np.float32).swapaxes(1, 0)
+
+        mb_dones = np.asarray(mb_dones, dtype=np.bool).swapaxes(1, 0)
+
+        mb_masks = mb_dones # Used for statefull models like LSTM's to mask state when done
+        mb_dones = mb_dones[:, 1:] # Used for calculating returns. The dones array is now aligned with rewards
+
+        # shapes are now [nenv, nsteps, []]
+        # When pulling from buffer, arrays will now be reshaped in place, preventing a deep copy.
+
+        return enc_obs, mb_obs, mb_actions, mb_rewards, mb_mus, mb_dones, mb_masks
+

baselines/acktr/README.md

@@ -0,0 +1,9 @@
+# ACKTR
+
+- Original paper: https://arxiv.org/abs/1708.05144
+- Baselines blog post: https://blog.openai.com/baselines-acktr-a2c/
+- `python -m baselines.run --alg=acktr --env=PongNoFrameskip-v4` runs the algorithm for 40M frames = 10M timesteps on an Atari Pong. See help (`-h`) for more options.
+- also refer to the repo-wide [README.md](../../README.md#training-models)
+
+## ACKTR with continuous action spaces
+The code of ACKTR has been refactored to handle both discrete and continuous action spaces uniformly. In the original version, discrete and continuous action spaces were handled by different code (actkr_disc.py and acktr_cont.py) with little overlap. If interested in the original version of the acktr for continuous action spaces, use `old_acktr_cont` branch. Note that original code performs better on the mujoco tasks than the refactored version; we are still investigating why. 

baselines/acktr/acktr.py

@@ -0,0 +1,158 @@
+import os.path as osp
+import time
+import functools
+import tensorflow as tf
+from baselines import logger
+
+from baselines.common import set_global_seeds, explained_variance
+from baselines.common.policies import build_policy
+from baselines.common.tf_util import get_session, save_variables, load_variables
+
+from baselines.a2c.runner import Runner
+from baselines.a2c.utils import Scheduler, find_trainable_variables
+from baselines.acktr import kfac
+from baselines.ppo2.ppo2 import safemean
+from collections import deque
+
+
+class Model(object):
+
+    def __init__(self, policy, ob_space, ac_space, nenvs,total_timesteps, nprocs=32, nsteps=20,
+                 ent_coef=0.01, vf_coef=0.5, vf_fisher_coef=1.0, lr=0.25, max_grad_norm=0.5,
+                 kfac_clip=0.001, lrschedule='linear', is_async=True):
+
+        self.sess = sess = get_session()
+        nbatch = nenvs * nsteps
+        with tf.variable_scope('acktr_model', reuse=tf.AUTO_REUSE):
+            self.model = step_model = policy(nenvs, 1, sess=sess)
+            self.model2 = train_model = policy(nenvs*nsteps, nsteps, sess=sess)
+
+        A = train_model.pdtype.sample_placeholder([None])
+        ADV = tf.placeholder(tf.float32, [nbatch])
+        R = tf.placeholder(tf.float32, [nbatch])
+        PG_LR = tf.placeholder(tf.float32, [])
+        VF_LR = tf.placeholder(tf.float32, [])
+
+        neglogpac = train_model.pd.neglogp(A)
+        self.logits = train_model.pi
+
+        ##training loss
+        pg_loss = tf.reduce_mean(ADV*neglogpac)
+        entropy = tf.reduce_mean(train_model.pd.entropy())
+        pg_loss = pg_loss - ent_coef * entropy
+        vf_loss = tf.losses.mean_squared_error(tf.squeeze(train_model.vf), R)
+        train_loss = pg_loss + vf_coef * vf_loss
+
+
+        ##Fisher loss construction
+        self.pg_fisher = pg_fisher_loss = -tf.reduce_mean(neglogpac)
+        sample_net = train_model.vf + tf.random_normal(tf.shape(train_model.vf))
+        self.vf_fisher = vf_fisher_loss = - vf_fisher_coef*tf.reduce_mean(tf.pow(train_model.vf - tf.stop_gradient(sample_net), 2))
+        self.joint_fisher = joint_fisher_loss = pg_fisher_loss + vf_fisher_loss
+
+        self.params=params = find_trainable_variables("acktr_model")
+
+        self.grads_check = grads = tf.gradients(train_loss,params)
+
+        with tf.device('/gpu:0'):
+            self.optim = optim = kfac.KfacOptimizer(learning_rate=PG_LR, clip_kl=kfac_clip,\
+                momentum=0.9, kfac_update=1, epsilon=0.01,\
+                stats_decay=0.99, is_async=is_async, cold_iter=10, max_grad_norm=max_grad_norm)
+
+            # update_stats_op = optim.compute_and_apply_stats(joint_fisher_loss, var_list=params)
+            optim.compute_and_apply_stats(joint_fisher_loss, var_list=params)
+            train_op, q_runner = optim.apply_gradients(list(zip(grads,params)))
+        self.q_runner = q_runner
+        self.lr = Scheduler(v=lr, nvalues=total_timesteps, schedule=lrschedule)
+
+        def train(obs, states, rewards, masks, actions, values):
+            advs = rewards - values
+            for step in range(len(obs)):
+                cur_lr = self.lr.value()
+
+            td_map = {train_model.X:obs, A:actions, ADV:advs, R:rewards, PG_LR:cur_lr, VF_LR:cur_lr}
+            if states is not None:
+                td_map[train_model.S] = states
+                td_map[train_model.M] = masks
+
+            policy_loss, value_loss, policy_entropy, _ = sess.run(
+                [pg_loss, vf_loss, entropy, train_op],
+                td_map
+            )
+            return policy_loss, value_loss, policy_entropy
+
+
+        self.train = train
+        self.save = functools.partial(save_variables, sess=sess)
+        self.load = functools.partial(load_variables, sess=sess)
+        self.train_model = train_model
+        self.step_model = step_model
+        self.step = step_model.step
+        self.value = step_model.value
+        self.initial_state = step_model.initial_state
+        tf.global_variables_initializer().run(session=sess)
+
+def learn(network, env, seed, total_timesteps=int(40e6), gamma=0.99, log_interval=100, nprocs=32, nsteps=20,
+                 ent_coef=0.01, vf_coef=0.5, vf_fisher_coef=1.0, lr=0.25, max_grad_norm=0.5,
+                 kfac_clip=0.001, save_interval=None, lrschedule='linear', load_path=None, is_async=True, **network_kwargs):
+    set_global_seeds(seed)
+
+
+    if network == 'cnn':
+        network_kwargs['one_dim_bias'] = True
+
+    policy = build_policy(env, network, **network_kwargs)
+
+    nenvs = env.num_envs
+    ob_space = env.observation_space
+    ac_space = env.action_space
+    make_model = lambda : Model(policy, ob_space, ac_space, nenvs, total_timesteps, nprocs=nprocs, nsteps
+                                =nsteps, ent_coef=ent_coef, vf_coef=vf_coef, vf_fisher_coef=
+                                vf_fisher_coef, lr=lr, max_grad_norm=max_grad_norm, kfac_clip=kfac_clip,
+                                lrschedule=lrschedule, is_async=is_async)
+    if save_interval and logger.get_dir():
+        import cloudpickle
+        with open(osp.join(logger.get_dir(), 'make_model.pkl'), 'wb') as fh:
+            fh.write(cloudpickle.dumps(make_model))
+    model = make_model()
+
+    if load_path is not None:
+        model.load(load_path)
+
+    runner = Runner(env, model, nsteps=nsteps, gamma=gamma)
+    epinfobuf = deque(maxlen=100)
+    nbatch = nenvs*nsteps
+    tstart = time.time()
+    coord = tf.train.Coordinator()
+    if is_async:
+        enqueue_threads = model.q_runner.create_threads(model.sess, coord=coord, start=True)
+    else:
+        enqueue_threads = []
+
+    for update in range(1, total_timesteps//nbatch+1):
+        obs, states, rewards, masks, actions, values, epinfos = runner.run()
+        epinfobuf.extend(epinfos)
+        policy_loss, value_loss, policy_entropy = model.train(obs, states, rewards, masks, actions, values)
+        model.old_obs = obs
+        nseconds = time.time()-tstart
+        fps = int((update*nbatch)/nseconds)
+        if update % log_interval == 0 or update == 1:
+            ev = explained_variance(values, rewards)
+            logger.record_tabular("nupdates", update)
+            logger.record_tabular("total_timesteps", update*nbatch)
+            logger.record_tabular("fps", fps)
+            logger.record_tabular("policy_entropy", float(policy_entropy))
+            logger.record_tabular("policy_loss", float(policy_loss))
+            logger.record_tabular("value_loss", float(value_loss))
+            logger.record_tabular("explained_variance", float(ev))
+            logger.record_tabular("eprewmean", safemean([epinfo['r'] for epinfo in epinfobuf]))
+            logger.record_tabular("eplenmean", safemean([epinfo['l'] for epinfo in epinfobuf]))
+            logger.dump_tabular()
+
+        if save_interval and (update % save_interval == 0 or update == 1) and logger.get_dir():
+            savepath = osp.join(logger.get_dir(), 'checkpoint%.5i'%update)
+            print('Saving to', savepath)
+            model.save(savepath)
+    coord.request_stop()
+    coord.join(enqueue_threads)
+    return model

baselines/acktr/defaults.py

@@ -0,0 +1,5 @@
+def mujoco():
+    return dict(
+        nsteps=2500,
+        value_network='copy'
+    )

baselines/acktr/kfac.py

@@ -0,0 +1,928 @@
+import tensorflow as tf
+import numpy as np
+import re
+
+ # flake8: noqa F403, F405
+from baselines.acktr.kfac_utils import *
+from functools import reduce
+
+KFAC_OPS = ['MatMul', 'Conv2D', 'BiasAdd']
+KFAC_DEBUG = False
+
+
+class KfacOptimizer():
+    # note that KfacOptimizer will be truly synchronous (and thus deterministic) only if a single-threaded session is used
+    def __init__(self, learning_rate=0.01, momentum=0.9, clip_kl=0.01, kfac_update=2, stats_accum_iter=60, full_stats_init=False, cold_iter=100, cold_lr=None, is_async=False, async_stats=False, epsilon=1e-2, stats_decay=0.95, blockdiag_bias=False, channel_fac=False, factored_damping=False, approxT2=False, use_float64=False, weight_decay_dict={},max_grad_norm=0.5):
+        self.max_grad_norm = max_grad_norm
+        self._lr = learning_rate
+        self._momentum = momentum
+        self._clip_kl = clip_kl
+        self._channel_fac = channel_fac
+        self._kfac_update = kfac_update
+        self._async = is_async
+        self._async_stats = async_stats
+        self._epsilon = epsilon
+        self._stats_decay = stats_decay
+        self._blockdiag_bias = blockdiag_bias
+        self._approxT2 = approxT2
+        self._use_float64 = use_float64
+        self._factored_damping = factored_damping
+        self._cold_iter = cold_iter
+        if cold_lr == None:
+            # good heuristics
+            self._cold_lr = self._lr# * 3.
+        else:
+            self._cold_lr = cold_lr
+        self._stats_accum_iter = stats_accum_iter
+        self._weight_decay_dict = weight_decay_dict
+        self._diag_init_coeff = 0.
+        self._full_stats_init = full_stats_init
+        if not self._full_stats_init:
+            self._stats_accum_iter = self._cold_iter
+
+        self.sgd_step = tf.Variable(0, name='KFAC/sgd_step', trainable=False)
+        self.global_step = tf.Variable(
+            0, name='KFAC/global_step', trainable=False)
+        self.cold_step = tf.Variable(0, name='KFAC/cold_step', trainable=False)
+        self.factor_step = tf.Variable(
+            0, name='KFAC/factor_step', trainable=False)
+        self.stats_step = tf.Variable(
+            0, name='KFAC/stats_step', trainable=False)
+        self.vFv = tf.Variable(0., name='KFAC/vFv', trainable=False)
+
+        self.factors = {}
+        self.param_vars = []
+        self.stats = {}
+        self.stats_eigen = {}
+
+    def getFactors(self, g, varlist):
+        graph = tf.get_default_graph()
+        factorTensors = {}
+        fpropTensors = []
+        bpropTensors = []
+        opTypes = []
+        fops = []
+
+        def searchFactors(gradient, graph):
+            # hard coded search stratergy
+            bpropOp = gradient.op
+            bpropOp_name = bpropOp.name
+
+            bTensors = []
+            fTensors = []
+
+            # combining additive gradient, assume they are the same op type and
+            # indepedent
+            if 'AddN' in bpropOp_name:
+                factors = []
+                for g in gradient.op.inputs:
+                    factors.append(searchFactors(g, graph))
+                op_names = [item['opName'] for item in factors]
+                # TO-DO: need to check all the attribute of the ops as well
+                print (gradient.name)
+                print (op_names)
+                print (len(np.unique(op_names)))
+                assert len(np.unique(op_names)) == 1, gradient.name + \
+                    ' is shared among different computation OPs'
+
+                bTensors = reduce(lambda x, y: x + y,
+                                  [item['bpropFactors'] for item in factors])
+                if len(factors[0]['fpropFactors']) > 0:
+                    fTensors = reduce(
+                        lambda x, y: x + y, [item['fpropFactors'] for item in factors])
+                fpropOp_name = op_names[0]
+                fpropOp = factors[0]['op']
+            else:
+                fpropOp_name = re.search(
+                    'gradientsSampled(_[0-9]+|)/(.+?)_grad', bpropOp_name).group(2)
+                fpropOp = graph.get_operation_by_name(fpropOp_name)
+                if fpropOp.op_def.name in KFAC_OPS:
+                    # Known OPs
+                    ###
+                    bTensor = [
+                        i for i in bpropOp.inputs if 'gradientsSampled' in i.name][-1]
+                    bTensorShape = fpropOp.outputs[0].get_shape()
+                    if bTensor.get_shape()[0].value == None:
+                        bTensor.set_shape(bTensorShape)
+                    bTensors.append(bTensor)
+                    ###
+                    if fpropOp.op_def.name == 'BiasAdd':
+                        fTensors = []
+                    else:
+                        fTensors.append(
+                            [i for i in fpropOp.inputs if param.op.name not in i.name][0])
+                    fpropOp_name = fpropOp.op_def.name
+                else:
+                    # unknown OPs, block approximation used
+                    bInputsList = [i for i in bpropOp.inputs[
+                        0].op.inputs if 'gradientsSampled' in i.name if 'Shape' not in i.name]
+                    if len(bInputsList) > 0:
+                        bTensor = bInputsList[0]
+                        bTensorShape = fpropOp.outputs[0].get_shape()
+                        if len(bTensor.get_shape()) > 0 and bTensor.get_shape()[0].value == None:
+                            bTensor.set_shape(bTensorShape)
+                        bTensors.append(bTensor)
+                    fpropOp_name = opTypes.append('UNK-' + fpropOp.op_def.name)
+
+            return {'opName': fpropOp_name, 'op': fpropOp, 'fpropFactors': fTensors, 'bpropFactors': bTensors}
+
+        for t, param in zip(g, varlist):
+            if KFAC_DEBUG:
+                print(('get factor for '+param.name))
+            factors = searchFactors(t, graph)
+            factorTensors[param] = factors
+
+        ########
+        # check associated weights and bias for homogeneous coordinate representation
+        # and check redundent factors
+        # TO-DO: there may be a bug to detect associate bias and weights for
+        # forking layer, e.g. in inception models.
+        for param in varlist:
+            factorTensors[param]['assnWeights'] = None
+            factorTensors[param]['assnBias'] = None
+        for param in varlist:
+            if factorTensors[param]['opName'] == 'BiasAdd':
+                factorTensors[param]['assnWeights'] = None
+                for item in varlist:
+                    if len(factorTensors[item]['bpropFactors']) > 0:
+                        if (set(factorTensors[item]['bpropFactors']) == set(factorTensors[param]['bpropFactors'])) and (len(factorTensors[item]['fpropFactors']) > 0):
+                            factorTensors[param]['assnWeights'] = item
+                            factorTensors[item]['assnBias'] = param
+                            factorTensors[param]['bpropFactors'] = factorTensors[
+                                item]['bpropFactors']
+
+        ########
+
+        ########
+        # concatenate the additive gradients along the batch dimension, i.e.
+        # assuming independence structure
+        for key in ['fpropFactors', 'bpropFactors']:
+            for i, param in enumerate(varlist):
+                if len(factorTensors[param][key]) > 0:
+                    if (key + '_concat') not in factorTensors[param]:
+                        name_scope = factorTensors[param][key][0].name.split(':')[
+                            0]
+                        with tf.name_scope(name_scope):
+                            factorTensors[param][
+                                key + '_concat'] = tf.concat(factorTensors[param][key], 0)
+                else:
+                    factorTensors[param][key + '_concat'] = None
+                for j, param2 in enumerate(varlist[(i + 1):]):
+                    if (len(factorTensors[param][key]) > 0) and (set(factorTensors[param2][key]) == set(factorTensors[param][key])):
+                        factorTensors[param2][key] = factorTensors[param][key]
+                        factorTensors[param2][
+                            key + '_concat'] = factorTensors[param][key + '_concat']
+        ########
+
+        if KFAC_DEBUG:
+            for items in zip(varlist, fpropTensors, bpropTensors, opTypes):
+                print((items[0].name, factorTensors[item]))
+        self.factors = factorTensors
+        return factorTensors
+
+    def getStats(self, factors, varlist):
+        if len(self.stats) == 0:
+            # initialize stats variables on CPU because eigen decomp is
+            # computed on CPU
+            with tf.device('/cpu'):
+                tmpStatsCache = {}
+
+                # search for tensor factors and
+                # use block diag approx for the bias units
+                for var in varlist:
+                    fpropFactor = factors[var]['fpropFactors_concat']
+                    bpropFactor = factors[var]['bpropFactors_concat']
+                    opType = factors[var]['opName']
+                    if opType == 'Conv2D':
+                        Kh = var.get_shape()[0]
+                        Kw = var.get_shape()[1]
+                        C = fpropFactor.get_shape()[-1]
+
+                        Oh = bpropFactor.get_shape()[1]
+                        Ow = bpropFactor.get_shape()[2]
+                        if Oh == 1 and Ow == 1 and self._channel_fac:
+                            # factorization along the channels do not support
+                            # homogeneous coordinate
+                            var_assnBias = factors[var]['assnBias']
+                            if var_assnBias:
+                                factors[var]['assnBias'] = None
+                                factors[var_assnBias]['assnWeights'] = None
+                ##
+
+                for var in varlist:
+                    fpropFactor = factors[var]['fpropFactors_concat']
+                    bpropFactor = factors[var]['bpropFactors_concat']
+                    opType = factors[var]['opName']
+                    self.stats[var] = {'opName': opType,
+                                       'fprop_concat_stats': [],
+                                       'bprop_concat_stats': [],
+                                       'assnWeights': factors[var]['assnWeights'],
+                                       'assnBias': factors[var]['assnBias'],
+                                       }
+                    if fpropFactor is not None:
+                        if fpropFactor not in tmpStatsCache:
+                            if opType == 'Conv2D':
+                                Kh = var.get_shape()[0]
+                                Kw = var.get_shape()[1]
+                                C = fpropFactor.get_shape()[-1]
+
+                                Oh = bpropFactor.get_shape()[1]
+                                Ow = bpropFactor.get_shape()[2]
+                                if Oh == 1 and Ow == 1 and self._channel_fac:
+                                    # factorization along the channels
+                                    # assume independence between input channels and spatial
+                                    # 2K-1 x 2K-1 covariance matrix and C x C covariance matrix
+                                    # factorization along the channels do not
+                                    # support homogeneous coordinate, assnBias
+                                    # is always None
+                                    fpropFactor2_size = Kh * Kw
+                                    slot_fpropFactor_stats2 = tf.Variable(tf.diag(tf.ones(
+                                        [fpropFactor2_size])) * self._diag_init_coeff, name='KFAC_STATS/' + fpropFactor.op.name, trainable=False)
+                                    self.stats[var]['fprop_concat_stats'].append(
+                                        slot_fpropFactor_stats2)
+
+                                    fpropFactor_size = C
+                                else:
+                                    # 2K-1 x 2K-1 x C x C covariance matrix
+                                    # assume BHWC
+                                    fpropFactor_size = Kh * Kw * C
+                            else:
+                                # D x D covariance matrix
+                                fpropFactor_size = fpropFactor.get_shape()[-1]
+
+                            # use homogeneous coordinate
+                            if not self._blockdiag_bias and self.stats[var]['assnBias']:
+                                fpropFactor_size += 1
+
+                            slot_fpropFactor_stats = tf.Variable(tf.diag(tf.ones(
+                                [fpropFactor_size])) * self._diag_init_coeff, name='KFAC_STATS/' + fpropFactor.op.name, trainable=False)
+                            self.stats[var]['fprop_concat_stats'].append(
+                                slot_fpropFactor_stats)
+                            if opType != 'Conv2D':
+                                tmpStatsCache[fpropFactor] = self.stats[
+                                    var]['fprop_concat_stats']
+                        else:
+                            self.stats[var][
+                                'fprop_concat_stats'] = tmpStatsCache[fpropFactor]
+
+                    if bpropFactor is not None:
+                        # no need to collect backward stats for bias vectors if
+                        # using homogeneous coordinates
+                        if not((not self._blockdiag_bias) and self.stats[var]['assnWeights']):
+                            if bpropFactor not in tmpStatsCache:
+                                slot_bpropFactor_stats = tf.Variable(tf.diag(tf.ones([bpropFactor.get_shape(
+                                )[-1]])) * self._diag_init_coeff, name='KFAC_STATS/' + bpropFactor.op.name, trainable=False)
+                                self.stats[var]['bprop_concat_stats'].append(
+                                    slot_bpropFactor_stats)
+                                tmpStatsCache[bpropFactor] = self.stats[
+                                    var]['bprop_concat_stats']
+                            else:
+                                self.stats[var][
+                                    'bprop_concat_stats'] = tmpStatsCache[bpropFactor]
+
+        return self.stats
+
+    def compute_and_apply_stats(self, loss_sampled, var_list=None):
+        varlist = var_list
+        if varlist is None:
+            varlist = tf.trainable_variables()
+
+        stats = self.compute_stats(loss_sampled, var_list=varlist)
+        return self.apply_stats(stats)
+
+    def compute_stats(self, loss_sampled, var_list=None):
+        varlist = var_list
+        if varlist is None:
+            varlist = tf.trainable_variables()
+
+        gs = tf.gradients(loss_sampled, varlist, name='gradientsSampled')
+        self.gs = gs
+        factors = self.getFactors(gs, varlist)
+        stats = self.getStats(factors, varlist)
+
+        updateOps = []
+        statsUpdates = {}
+        statsUpdates_cache = {}
+        for var in varlist:
+            opType = factors[var]['opName']
+            fops = factors[var]['op']
+            fpropFactor = factors[var]['fpropFactors_concat']
+            fpropStats_vars = stats[var]['fprop_concat_stats']
+            bpropFactor = factors[var]['bpropFactors_concat']
+            bpropStats_vars = stats[var]['bprop_concat_stats']
+            SVD_factors = {}
+            for stats_var in fpropStats_vars:
+                stats_var_dim = int(stats_var.get_shape()[0])
+                if stats_var not in statsUpdates_cache:
+                    old_fpropFactor = fpropFactor
+                    B = (tf.shape(fpropFactor)[0])  # batch size
+                    if opType == 'Conv2D':
+                        strides = fops.get_attr("strides")
+                        padding = fops.get_attr("padding")
+                        convkernel_size = var.get_shape()[0:3]
+
+                        KH = int(convkernel_size[0])
+                        KW = int(convkernel_size[1])
+                        C = int(convkernel_size[2])
+                        flatten_size = int(KH * KW * C)
+
+                        Oh = int(bpropFactor.get_shape()[1])
+                        Ow = int(bpropFactor.get_shape()[2])
+
+                        if Oh == 1 and Ow == 1 and self._channel_fac:
+                                # factorization along the channels
+                                # assume independence among input channels
+                                # factor = B x 1 x 1 x (KH xKW x C)
+                                # patches = B x Oh x Ow x (KH xKW x C)
+                            if len(SVD_factors) == 0:
+                                if KFAC_DEBUG:
+                                    print(('approx %s act factor with rank-1 SVD factors' % (var.name)))
+                                # find closest rank-1 approx to the feature map
+                                S, U, V = tf.batch_svd(tf.reshape(
+                                    fpropFactor, [-1, KH * KW, C]))
+                                # get rank-1 approx slides
+                                sqrtS1 = tf.expand_dims(tf.sqrt(S[:, 0, 0]), 1)
+                                patches_k = U[:, :, 0] * sqrtS1  # B x KH*KW
+                                full_factor_shape = fpropFactor.get_shape()
+                                patches_k.set_shape(
+                                    [full_factor_shape[0], KH * KW])
+                                patches_c = V[:, :, 0] * sqrtS1  # B x C
+                                patches_c.set_shape([full_factor_shape[0], C])
+                                SVD_factors[C] = patches_c
+                                SVD_factors[KH * KW] = patches_k
+                            fpropFactor = SVD_factors[stats_var_dim]
+
+                        else:
+                            # poor mem usage implementation
+                            patches = tf.extract_image_patches(fpropFactor, ksizes=[1, convkernel_size[
+                                                               0], convkernel_size[1], 1], strides=strides, rates=[1, 1, 1, 1], padding=padding)
+
+                            if self._approxT2:
+                                if KFAC_DEBUG:
+                                    print(('approxT2 act fisher for %s' % (var.name)))
+                                # T^2 terms * 1/T^2, size: B x C
+                                fpropFactor = tf.reduce_mean(patches, [1, 2])
+                            else:
+                                # size: (B x Oh x Ow) x C
+                                fpropFactor = tf.reshape(
+                                    patches, [-1, flatten_size]) / Oh / Ow
+                    fpropFactor_size = int(fpropFactor.get_shape()[-1])
+                    if stats_var_dim == (fpropFactor_size + 1) and not self._blockdiag_bias:
+                        if opType == 'Conv2D' and not self._approxT2:
+                            # correct padding for numerical stability (we
+                            # divided out OhxOw from activations for T1 approx)
+                            fpropFactor = tf.concat([fpropFactor, tf.ones(
+                                [tf.shape(fpropFactor)[0], 1]) / Oh / Ow], 1)
+                        else:
+                            # use homogeneous coordinates
+                            fpropFactor = tf.concat(
+                                [fpropFactor, tf.ones([tf.shape(fpropFactor)[0], 1])], 1)
+
+                    # average over the number of data points in a batch
+                    # divided by B
+                    cov = tf.matmul(fpropFactor, fpropFactor,
+                                    transpose_a=True) / tf.cast(B, tf.float32)
+                    updateOps.append(cov)
+                    statsUpdates[stats_var] = cov
+                    if opType != 'Conv2D':
+                        # HACK: for convolution we recompute fprop stats for
+                        # every layer including forking layers
+                        statsUpdates_cache[stats_var] = cov
+
+            for stats_var in bpropStats_vars:
+                stats_var_dim = int(stats_var.get_shape()[0])
+                if stats_var not in statsUpdates_cache:
+                    old_bpropFactor = bpropFactor
+                    bpropFactor_shape = bpropFactor.get_shape()
+                    B = tf.shape(bpropFactor)[0]  # batch size
+                    C = int(bpropFactor_shape[-1])  # num channels
+                    if opType == 'Conv2D' or len(bpropFactor_shape) == 4:
+                        if fpropFactor is not None:
+                            if self._approxT2:
+                                if KFAC_DEBUG:
+                                    print(('approxT2 grad fisher for %s' % (var.name)))
+                                bpropFactor = tf.reduce_sum(
+                                    bpropFactor, [1, 2])  # T^2 terms * 1/T^2
+                            else:
+                                bpropFactor = tf.reshape(
+                                    bpropFactor, [-1, C]) * Oh * Ow  # T * 1/T terms
+                        else:
+                            # just doing block diag approx. spatial independent
+                            # structure does not apply here. summing over
+                            # spatial locations
+                            if KFAC_DEBUG:
+                                print(('block diag approx fisher for %s' % (var.name)))
+                            bpropFactor = tf.reduce_sum(bpropFactor, [1, 2])
+
+                    # assume sampled loss is averaged. TO-DO:figure out better
+                    # way to handle this
+                    bpropFactor *= tf.to_float(B)
+                    ##
+
+                    cov_b = tf.matmul(
+                        bpropFactor, bpropFactor, transpose_a=True) / tf.to_float(tf.shape(bpropFactor)[0])
+
+                    updateOps.append(cov_b)
+                    statsUpdates[stats_var] = cov_b
+                    statsUpdates_cache[stats_var] = cov_b
+
+        if KFAC_DEBUG:
+            aKey = list(statsUpdates.keys())[0]
+            statsUpdates[aKey] = tf.Print(statsUpdates[aKey],
+                                          [tf.convert_to_tensor('step:'),
+                                           self.global_step,
+                                           tf.convert_to_tensor(
+                                               'computing stats'),
+                                           ])
+        self.statsUpdates = statsUpdates
+        return statsUpdates
+
+    def apply_stats(self, statsUpdates):
+        """ compute stats and update/apply the new stats to the running average
+        """
+
+        def updateAccumStats():
+            if self._full_stats_init:
+                return tf.cond(tf.greater(self.sgd_step, self._cold_iter), lambda: tf.group(*self._apply_stats(statsUpdates, accumulate=True, accumulateCoeff=1. / self._stats_accum_iter)), tf.no_op)
+            else:
+                return tf.group(*self._apply_stats(statsUpdates, accumulate=True, accumulateCoeff=1. / self._stats_accum_iter))
+
+        def updateRunningAvgStats(statsUpdates, fac_iter=1):
+            # return tf.cond(tf.greater_equal(self.factor_step,
+            # tf.convert_to_tensor(fac_iter)), lambda:
+            # tf.group(*self._apply_stats(stats_list, varlist)), tf.no_op)
+            return tf.group(*self._apply_stats(statsUpdates))
+
+        if self._async_stats:
+            # asynchronous stats update
+            update_stats = self._apply_stats(statsUpdates)
+
+            queue = tf.FIFOQueue(1, [item.dtype for item in update_stats], shapes=[
+                                 item.get_shape() for item in update_stats])
+            enqueue_op = queue.enqueue(update_stats)
+
+            def dequeue_stats_op():
+                return queue.dequeue()
+            self.qr_stats = tf.train.QueueRunner(queue, [enqueue_op])
+            update_stats_op = tf.cond(tf.equal(queue.size(), tf.convert_to_tensor(
+                0)), tf.no_op, lambda: tf.group(*[dequeue_stats_op(), ]))
+        else:
+            # synchronous stats update
+            update_stats_op = tf.cond(tf.greater_equal(
+                self.stats_step, self._stats_accum_iter), lambda: updateRunningAvgStats(statsUpdates), updateAccumStats)
+        self._update_stats_op = update_stats_op
+        return update_stats_op
+
+    def _apply_stats(self, statsUpdates, accumulate=False, accumulateCoeff=0.):
+        updateOps = []
+        # obtain the stats var list
+        for stats_var in statsUpdates:
+            stats_new = statsUpdates[stats_var]
+            if accumulate:
+                # simple superbatch averaging
+                update_op = tf.assign_add(
+                    stats_var, accumulateCoeff * stats_new, use_locking=True)
+            else:
+                # exponential running averaging
+                update_op = tf.assign(
+                    stats_var, stats_var * self._stats_decay, use_locking=True)
+                update_op = tf.assign_add(
+                    update_op, (1. - self._stats_decay) * stats_new, use_locking=True)
+            updateOps.append(update_op)
+
+        with tf.control_dependencies(updateOps):
+            stats_step_op = tf.assign_add(self.stats_step, 1)
+
+        if KFAC_DEBUG:
+            stats_step_op = (tf.Print(stats_step_op,
+                                      [tf.convert_to_tensor('step:'),
+                                       self.global_step,
+                                       tf.convert_to_tensor('fac step:'),
+                                       self.factor_step,
+                                       tf.convert_to_tensor('sgd step:'),
+                                       self.sgd_step,
+                                       tf.convert_to_tensor('Accum:'),
+                                       tf.convert_to_tensor(accumulate),
+                                       tf.convert_to_tensor('Accum coeff:'),
+                                       tf.convert_to_tensor(accumulateCoeff),
+                                       tf.convert_to_tensor('stat step:'),
+                                       self.stats_step, updateOps[0], updateOps[1]]))
+        return [stats_step_op, ]
+
+    def getStatsEigen(self, stats=None):
+        if len(self.stats_eigen) == 0:
+            stats_eigen = {}
+            if stats is None:
+                stats = self.stats
+
+            tmpEigenCache = {}
+            with tf.device('/cpu:0'):
+                for var in stats:
+                    for key in ['fprop_concat_stats', 'bprop_concat_stats']:
+                        for stats_var in stats[var][key]:
+                            if stats_var not in tmpEigenCache:
+                                stats_dim = stats_var.get_shape()[1].value
+                                e = tf.Variable(tf.ones(
+                                    [stats_dim]), name='KFAC_FAC/' + stats_var.name.split(':')[0] + '/e', trainable=False)
+                                Q = tf.Variable(tf.diag(tf.ones(
+                                    [stats_dim])), name='KFAC_FAC/' + stats_var.name.split(':')[0] + '/Q', trainable=False)
+                                stats_eigen[stats_var] = {'e': e, 'Q': Q}
+                                tmpEigenCache[
+                                    stats_var] = stats_eigen[stats_var]
+                            else:
+                                stats_eigen[stats_var] = tmpEigenCache[
+                                    stats_var]
+            self.stats_eigen = stats_eigen
+        return self.stats_eigen
+
+    def computeStatsEigen(self):
+        """ compute the eigen decomp using copied var stats to avoid concurrent read/write from other queue """
+        # TO-DO: figure out why this op has delays (possibly moving
+        # eigenvectors around?)
+        with tf.device('/cpu:0'):
+            def removeNone(tensor_list):
+                local_list = []
+                for item in tensor_list:
+                    if item is not None:
+                        local_list.append(item)
+                return local_list
+
+            def copyStats(var_list):
+                print("copying stats to buffer tensors before eigen decomp")
+                redundant_stats = {}
+                copied_list = []
+                for item in var_list:
+                    if item is not None:
+                        if item not in redundant_stats:
+                            if self._use_float64:
+                                redundant_stats[item] = tf.cast(
+                                    tf.identity(item), tf.float64)
+                            else:
+                                redundant_stats[item] = tf.identity(item)
+                        copied_list.append(redundant_stats[item])
+                    else:
+                        copied_list.append(None)
+                return copied_list
+            #stats = [copyStats(self.fStats), copyStats(self.bStats)]
+            #stats = [self.fStats, self.bStats]
+
+            stats_eigen = self.stats_eigen
+            computedEigen = {}
+            eigen_reverse_lookup = {}
+            updateOps = []
+            # sync copied stats
+            # with tf.control_dependencies(removeNone(stats[0]) +
+            # removeNone(stats[1])):
+            with tf.control_dependencies([]):
+                for stats_var in stats_eigen:
+                    if stats_var not in computedEigen:
+                        eigens = tf.self_adjoint_eig(stats_var)
+                        e = eigens[0]
+                        Q = eigens[1]
+                        if self._use_float64:
+                            e = tf.cast(e, tf.float32)
+                            Q = tf.cast(Q, tf.float32)
+                        updateOps.append(e)
+                        updateOps.append(Q)
+                        computedEigen[stats_var] = {'e': e, 'Q': Q}
+                        eigen_reverse_lookup[e] = stats_eigen[stats_var]['e']
+                        eigen_reverse_lookup[Q] = stats_eigen[stats_var]['Q']
+
+            self.eigen_reverse_lookup = eigen_reverse_lookup
+            self.eigen_update_list = updateOps
+
+            if KFAC_DEBUG:
+                self.eigen_update_list = [item for item in updateOps]
+                with tf.control_dependencies(updateOps):
+                    updateOps.append(tf.Print(tf.constant(
+                        0.), [tf.convert_to_tensor('computed factor eigen')]))
+
+        return updateOps
+
+    def applyStatsEigen(self, eigen_list):
+        updateOps = []
+        print(('updating %d eigenvalue/vectors' % len(eigen_list)))
+        for i, (tensor, mark) in enumerate(zip(eigen_list, self.eigen_update_list)):
+            stats_eigen_var = self.eigen_reverse_lookup[mark]
+            updateOps.append(
+                tf.assign(stats_eigen_var, tensor, use_locking=True))
+
+        with tf.control_dependencies(updateOps):
+            factor_step_op = tf.assign_add(self.factor_step, 1)
+            updateOps.append(factor_step_op)
+            if KFAC_DEBUG:
+                updateOps.append(tf.Print(tf.constant(
+                    0.), [tf.convert_to_tensor('updated kfac factors')]))
+        return updateOps
+
+    def getKfacPrecondUpdates(self, gradlist, varlist):
+        updatelist = []
+        vg = 0.
+
+        assert len(self.stats) > 0
+        assert len(self.stats_eigen) > 0
+        assert len(self.factors) > 0
+        counter = 0
+
+        grad_dict = {var: grad for grad, var in zip(gradlist, varlist)}
+
+        for grad, var in zip(gradlist, varlist):
+            GRAD_RESHAPE = False
+            GRAD_TRANSPOSE = False
+
+            fpropFactoredFishers = self.stats[var]['fprop_concat_stats']
+            bpropFactoredFishers = self.stats[var]['bprop_concat_stats']
+
+            if (len(fpropFactoredFishers) + len(bpropFactoredFishers)) > 0:
+                counter += 1
+                GRAD_SHAPE = grad.get_shape()
+                if len(grad.get_shape()) > 2:
+                    # reshape conv kernel parameters
+                    KW = int(grad.get_shape()[0])
+                    KH = int(grad.get_shape()[1])
+                    C = int(grad.get_shape()[2])
+                    D = int(grad.get_shape()[3])
+
+                    if len(fpropFactoredFishers) > 1 and self._channel_fac:
+                        # reshape conv kernel parameters into tensor
+                        grad = tf.reshape(grad, [KW * KH, C, D])
+                    else:
+                        # reshape conv kernel parameters into 2D grad
+                        grad = tf.reshape(grad, [-1, D])
+                    GRAD_RESHAPE = True
+                elif len(grad.get_shape()) == 1:
+                    # reshape bias or 1D parameters
+                    D = int(grad.get_shape()[0])
+
+                    grad = tf.expand_dims(grad, 0)
+                    GRAD_RESHAPE = True
+                else:
+                    # 2D parameters
+                    C = int(grad.get_shape()[0])
+                    D = int(grad.get_shape()[1])
+
+                if (self.stats[var]['assnBias'] is not None) and not self._blockdiag_bias:
+                    # use homogeneous coordinates only works for 2D grad.
+                    # TO-DO: figure out how to factorize bias grad
+                    # stack bias grad
+                    var_assnBias = self.stats[var]['assnBias']
+                    grad = tf.concat(
+                        [grad, tf.expand_dims(grad_dict[var_assnBias], 0)], 0)
+
+                # project gradient to eigen space and reshape the eigenvalues
+                # for broadcasting
+                eigVals = []
+
+                for idx, stats in enumerate(self.stats[var]['fprop_concat_stats']):
+                    Q = self.stats_eigen[stats]['Q']
+                    e = detectMinVal(self.stats_eigen[stats][
+                                     'e'], var, name='act', debug=KFAC_DEBUG)
+
+                    Q, e = factorReshape(Q, e, grad, facIndx=idx, ftype='act')
+                    eigVals.append(e)
+                    grad = gmatmul(Q, grad, transpose_a=True, reduce_dim=idx)
+
+                for idx, stats in enumerate(self.stats[var]['bprop_concat_stats']):
+                    Q = self.stats_eigen[stats]['Q']
+                    e = detectMinVal(self.stats_eigen[stats][
+                                     'e'], var, name='grad', debug=KFAC_DEBUG)
+
+                    Q, e = factorReshape(Q, e, grad, facIndx=idx, ftype='grad')
+                    eigVals.append(e)
+                    grad = gmatmul(grad, Q, transpose_b=False, reduce_dim=idx)
+                ##
+
+                #####
+                # whiten using eigenvalues
+                weightDecayCoeff = 0.
+                if var in self._weight_decay_dict:
+                    weightDecayCoeff = self._weight_decay_dict[var]
+                    if KFAC_DEBUG:
+                        print(('weight decay coeff for %s is %f' % (var.name, weightDecayCoeff)))
+
+                if self._factored_damping:
+                    if KFAC_DEBUG:
+                        print(('use factored damping for %s' % (var.name)))
+                    coeffs = 1.
+                    num_factors = len(eigVals)
+                    # compute the ratio of two trace norm of the left and right
+                    # KFac matrices, and their generalization
+                    if len(eigVals) == 1:
+                        damping = self._epsilon + weightDecayCoeff
+                    else:
+                        damping = tf.pow(
+                            self._epsilon + weightDecayCoeff, 1. / num_factors)
+                    eigVals_tnorm_avg = [tf.reduce_mean(
+                        tf.abs(e)) for e in eigVals]
+                    for e, e_tnorm in zip(eigVals, eigVals_tnorm_avg):
+                        eig_tnorm_negList = [
+                            item for item in eigVals_tnorm_avg if item != e_tnorm]
+                        if len(eigVals) == 1:
+                            adjustment = 1.
+                        elif len(eigVals) == 2:
+                            adjustment = tf.sqrt(
+                                e_tnorm / eig_tnorm_negList[0])
+                        else:
+                            eig_tnorm_negList_prod = reduce(
+                                lambda x, y: x * y, eig_tnorm_negList)
+                            adjustment = tf.pow(
+                                tf.pow(e_tnorm, num_factors - 1.) / eig_tnorm_negList_prod, 1. / num_factors)
+                        coeffs *= (e + adjustment * damping)
+                else:
+                    coeffs = 1.
+                    damping = (self._epsilon + weightDecayCoeff)
+                    for e in eigVals:
+                        coeffs *= e
+                    coeffs += damping
+
+                #grad = tf.Print(grad, [tf.convert_to_tensor('1'), tf.convert_to_tensor(var.name), grad.get_shape()])
+
+                grad /= coeffs
+
+                #grad = tf.Print(grad, [tf.convert_to_tensor('2'), tf.convert_to_tensor(var.name), grad.get_shape()])
+                #####
+                # project gradient back to euclidean space
+                for idx, stats in enumerate(self.stats[var]['fprop_concat_stats']):
+                    Q = self.stats_eigen[stats]['Q']
+                    grad = gmatmul(Q, grad, transpose_a=False, reduce_dim=idx)
+
+                for idx, stats in enumerate(self.stats[var]['bprop_concat_stats']):
+                    Q = self.stats_eigen[stats]['Q']
+                    grad = gmatmul(grad, Q, transpose_b=True, reduce_dim=idx)
+                ##
+
+                #grad = tf.Print(grad, [tf.convert_to_tensor('3'), tf.convert_to_tensor(var.name), grad.get_shape()])
+                if (self.stats[var]['assnBias'] is not None) and not self._blockdiag_bias:
+                    # use homogeneous coordinates only works for 2D grad.
+                    # TO-DO: figure out how to factorize bias grad
+                    # un-stack bias grad
+                    var_assnBias = self.stats[var]['assnBias']
+                    C_plus_one = int(grad.get_shape()[0])
+                    grad_assnBias = tf.reshape(tf.slice(grad,
+                                                        begin=[
+                                                            C_plus_one - 1, 0],
+                                                        size=[1, -1]), var_assnBias.get_shape())
+                    grad_assnWeights = tf.slice(grad,
+                                                begin=[0, 0],
+                                                size=[C_plus_one - 1, -1])
+                    grad_dict[var_assnBias] = grad_assnBias
+                    grad = grad_assnWeights
+
+                #grad = tf.Print(grad, [tf.convert_to_tensor('4'), tf.convert_to_tensor(var.name), grad.get_shape()])
+                if GRAD_RESHAPE:
+                    grad = tf.reshape(grad, GRAD_SHAPE)
+
+                grad_dict[var] = grad
+
+        print(('projecting %d gradient matrices' % counter))
+
+        for g, var in zip(gradlist, varlist):
+            grad = grad_dict[var]
+            ### clipping ###
+            if KFAC_DEBUG:
+                print(('apply clipping to %s' % (var.name)))
+            tf.Print(grad, [tf.sqrt(tf.reduce_sum(tf.pow(grad, 2)))], "Euclidean norm of new grad")
+            local_vg = tf.reduce_sum(grad * g * (self._lr * self._lr))
+            vg += local_vg
+
+        # recale everything
+        if KFAC_DEBUG:
+            print('apply vFv clipping')
+
+        scaling = tf.minimum(1., tf.sqrt(self._clip_kl / vg))
+        if KFAC_DEBUG:
+            scaling = tf.Print(scaling, [tf.convert_to_tensor(
+                'clip: '), scaling, tf.convert_to_tensor(' vFv: '), vg])
+        with tf.control_dependencies([tf.assign(self.vFv, vg)]):
+            updatelist = [grad_dict[var] for var in varlist]
+            for i, item in enumerate(updatelist):
+                updatelist[i] = scaling * item
+
+        return updatelist
+
+    def compute_gradients(self, loss, var_list=None):
+        varlist = var_list
+        if varlist is None:
+            varlist = tf.trainable_variables()
+        g = tf.gradients(loss, varlist)
+
+        return [(a, b) for a, b in zip(g, varlist)]
+
+    def apply_gradients_kfac(self, grads):
+        g, varlist = list(zip(*grads))
+
+        if len(self.stats_eigen) == 0:
+            self.getStatsEigen()
+
+        qr = None
+        # launch eigen-decomp on a queue thread
+        if self._async:
+            print('Use async eigen decomp')
+            # get a list of factor loading tensors
+            factorOps_dummy = self.computeStatsEigen()
+
+            # define a queue for the list of factor loading tensors
+            queue = tf.FIFOQueue(1, [item.dtype for item in factorOps_dummy], shapes=[
+                                 item.get_shape() for item in factorOps_dummy])
+            enqueue_op = tf.cond(tf.logical_and(tf.equal(tf.mod(self.stats_step, self._kfac_update), tf.convert_to_tensor(
+                0)), tf.greater_equal(self.stats_step, self._stats_accum_iter)), lambda: queue.enqueue(self.computeStatsEigen()), tf.no_op)
+
+            def dequeue_op():
+                return queue.dequeue()
+
+            qr = tf.train.QueueRunner(queue, [enqueue_op])
+
+        updateOps = []
+        global_step_op = tf.assign_add(self.global_step, 1)
+        updateOps.append(global_step_op)
+
+        with tf.control_dependencies([global_step_op]):
+
+            # compute updates
+            assert self._update_stats_op != None
+            updateOps.append(self._update_stats_op)
+            dependency_list = []
+            if not self._async:
+                dependency_list.append(self._update_stats_op)
+
+            with tf.control_dependencies(dependency_list):
+                def no_op_wrapper():
+                    return tf.group(*[tf.assign_add(self.cold_step, 1)])
+
+                if not self._async:
+                    # synchronous eigen-decomp updates
+                    updateFactorOps = tf.cond(tf.logical_and(tf.equal(tf.mod(self.stats_step, self._kfac_update),
+                                                                      tf.convert_to_tensor(0)),
+                                                             tf.greater_equal(self.stats_step, self._stats_accum_iter)), lambda: tf.group(*self.applyStatsEigen(self.computeStatsEigen())), no_op_wrapper)
+                else:
+                    # asynchronous eigen-decomp updates using queue
+                    updateFactorOps = tf.cond(tf.greater_equal(self.stats_step, self._stats_accum_iter),
+                                              lambda: tf.cond(tf.equal(queue.size(), tf.convert_to_tensor(0)),
+                                                              tf.no_op,
+
+                                                              lambda: tf.group(
+                                                                  *self.applyStatsEigen(dequeue_op())),
+                                                              ),
+                                              no_op_wrapper)
+
+                updateOps.append(updateFactorOps)
+
+                with tf.control_dependencies([updateFactorOps]):
+                    def gradOp():
+                        return list(g)
+
+                    def getKfacGradOp():
+                        return self.getKfacPrecondUpdates(g, varlist)
+                    u = tf.cond(tf.greater(self.factor_step,
+                                           tf.convert_to_tensor(0)), getKfacGradOp, gradOp)
+
+                    optim = tf.train.MomentumOptimizer(
+                        self._lr * (1. - self._momentum), self._momentum)
+                    #optim = tf.train.AdamOptimizer(self._lr, epsilon=0.01)
+
+                    def optimOp():
+                        def updateOptimOp():
+                            if self._full_stats_init:
+                                return tf.cond(tf.greater(self.factor_step, tf.convert_to_tensor(0)), lambda: optim.apply_gradients(list(zip(u, varlist))), tf.no_op)
+                            else:
+                                return optim.apply_gradients(list(zip(u, varlist)))
+                        if self._full_stats_init:
+                            return tf.cond(tf.greater_equal(self.stats_step, self._stats_accum_iter), updateOptimOp, tf.no_op)
+                        else:
+                            return tf.cond(tf.greater_equal(self.sgd_step, self._cold_iter), updateOptimOp, tf.no_op)
+                    updateOps.append(optimOp())
+
+        return tf.group(*updateOps), qr
+
+    def apply_gradients(self, grads):
+        coldOptim = tf.train.MomentumOptimizer(
+            self._cold_lr, self._momentum)
+
+        def coldSGDstart():
+            sgd_grads, sgd_var = zip(*grads)
+
+            if self.max_grad_norm != None:
+                sgd_grads, sgd_grad_norm = tf.clip_by_global_norm(sgd_grads,self.max_grad_norm)
+
+            sgd_grads = list(zip(sgd_grads,sgd_var))
+
+            sgd_step_op = tf.assign_add(self.sgd_step, 1)
+            coldOptim_op = coldOptim.apply_gradients(sgd_grads)
+            if KFAC_DEBUG:
+                with tf.control_dependencies([sgd_step_op, coldOptim_op]):
+                    sgd_step_op = tf.Print(
+                        sgd_step_op, [self.sgd_step, tf.convert_to_tensor('doing cold sgd step')])
+            return tf.group(*[sgd_step_op, coldOptim_op])
+
+        kfacOptim_op, qr = self.apply_gradients_kfac(grads)
+
+        def warmKFACstart():
+            return kfacOptim_op
+
+        return tf.cond(tf.greater(self.sgd_step, self._cold_iter), warmKFACstart, coldSGDstart), qr
+
+    def minimize(self, loss, loss_sampled, var_list=None):
+        grads = self.compute_gradients(loss, var_list=var_list)
+        update_stats_op = self.compute_and_apply_stats(
+            loss_sampled, var_list=var_list)
+        return self.apply_gradients(grads)

baselines/acktr/kfac_utils.py

@@ -0,0 +1,86 @@
+import tensorflow as tf
+
+def gmatmul(a, b, transpose_a=False, transpose_b=False, reduce_dim=None):
+    assert reduce_dim is not None
+
+    # weird batch matmul
+    if len(a.get_shape()) == 2 and len(b.get_shape()) > 2:
+        # reshape reduce_dim to the left most dim in b
+        b_shape = b.get_shape()
+        if reduce_dim != 0:
+            b_dims = list(range(len(b_shape)))
+            b_dims.remove(reduce_dim)
+            b_dims.insert(0, reduce_dim)
+            b = tf.transpose(b, b_dims)
+        b_t_shape = b.get_shape()
+        b = tf.reshape(b, [int(b_shape[reduce_dim]), -1])
+        result = tf.matmul(a, b, transpose_a=transpose_a,
+                           transpose_b=transpose_b)
+        result = tf.reshape(result, b_t_shape)
+        if reduce_dim != 0:
+            b_dims = list(range(len(b_shape)))
+            b_dims.remove(0)
+            b_dims.insert(reduce_dim, 0)
+            result = tf.transpose(result, b_dims)
+        return result
+
+    elif len(a.get_shape()) > 2 and len(b.get_shape()) == 2:
+        # reshape reduce_dim to the right most dim in a
+        a_shape = a.get_shape()
+        outter_dim = len(a_shape) - 1
+        reduce_dim = len(a_shape) - reduce_dim - 1
+        if reduce_dim != outter_dim:
+            a_dims = list(range(len(a_shape)))
+            a_dims.remove(reduce_dim)
+            a_dims.insert(outter_dim, reduce_dim)
+            a = tf.transpose(a, a_dims)
+        a_t_shape = a.get_shape()
+        a = tf.reshape(a, [-1, int(a_shape[reduce_dim])])
+        result = tf.matmul(a, b, transpose_a=transpose_a,
+                           transpose_b=transpose_b)
+        result = tf.reshape(result, a_t_shape)
+        if reduce_dim != outter_dim:
+            a_dims = list(range(len(a_shape)))
+            a_dims.remove(outter_dim)
+            a_dims.insert(reduce_dim, outter_dim)
+            result = tf.transpose(result, a_dims)
+        return result
+
+    elif len(a.get_shape()) == 2 and len(b.get_shape()) == 2:
+        return tf.matmul(a, b, transpose_a=transpose_a, transpose_b=transpose_b)
+
+    assert False, 'something went wrong'
+
+
+def clipoutNeg(vec, threshold=1e-6):
+    mask = tf.cast(vec > threshold, tf.float32)
+    return mask * vec
+
+
+def detectMinVal(input_mat, var, threshold=1e-6, name='', debug=False):
+    eigen_min = tf.reduce_min(input_mat)
+    eigen_max = tf.reduce_max(input_mat)
+    eigen_ratio = eigen_max / eigen_min
+    input_mat_clipped = clipoutNeg(input_mat, threshold)
+
+    if debug:
+        input_mat_clipped = tf.cond(tf.logical_or(tf.greater(eigen_ratio, 0.), tf.less(eigen_ratio, -500)), lambda: input_mat_clipped, lambda: tf.Print(
+            input_mat_clipped, [tf.convert_to_tensor('screwed ratio ' + name + ' eigen values!!!'), tf.convert_to_tensor(var.name), eigen_min, eigen_max, eigen_ratio]))
+
+    return input_mat_clipped
+
+
+def factorReshape(Q, e, grad, facIndx=0, ftype='act'):
+    grad_shape = grad.get_shape()
+    if ftype == 'act':
+        assert e.get_shape()[0] == grad_shape[facIndx]
+        expanded_shape = [1, ] * len(grad_shape)
+        expanded_shape[facIndx] = -1
+        e = tf.reshape(e, expanded_shape)
+    if ftype == 'grad':
+        assert e.get_shape()[0] == grad_shape[len(grad_shape) - facIndx - 1]
+        expanded_shape = [1, ] * len(grad_shape)
+        expanded_shape[len(grad_shape) - facIndx - 1] = -1
+        e = tf.reshape(e, expanded_shape)
+
+    return Q, e

baselines/acktr/utils.py

@@ -0,0 +1,28 @@
+import tensorflow as tf
+
+def dense(x, size, name, weight_init=None, bias_init=0, weight_loss_dict=None, reuse=None):
+    with tf.variable_scope(name, reuse=reuse):
+        assert (len(tf.get_variable_scope().name.split('/')) == 2)
+
+        w = tf.get_variable("w", [x.get_shape()[1], size], initializer=weight_init)
+        b = tf.get_variable("b", [size], initializer=tf.constant_initializer(bias_init))
+        weight_decay_fc = 3e-4
+
+        if weight_loss_dict is not None:
+            weight_decay = tf.multiply(tf.nn.l2_loss(w), weight_decay_fc, name='weight_decay_loss')
+            if weight_loss_dict is not None:
+                weight_loss_dict[w] = weight_decay_fc
+                weight_loss_dict[b] = 0.0
+
+            tf.add_to_collection(tf.get_variable_scope().name.split('/')[0] + '_' + 'losses', weight_decay)
+
+        return tf.nn.bias_add(tf.matmul(x, w), b)
+
+def kl_div(action_dist1, action_dist2, action_size):
+    mean1, std1 = action_dist1[:, :action_size], action_dist1[:, action_size:]
+    mean2, std2 = action_dist2[:, :action_size], action_dist2[:, action_size:]
+
+    numerator = tf.square(mean1 - mean2) + tf.square(std1) - tf.square(std2)
+    denominator = 2 * tf.square(std2) + 1e-8
+    return tf.reduce_sum(
+        numerator/denominator + tf.log(std2) - tf.log(std1),reduction_indices=-1)

baselines/bench/__init__.py

@@ -1,2 +1,2 @@
-from baselines.bench.benchmarks import * # noqa: F403 F401
-from baselines.bench.monitor import * # noqa: F403 F401
+from baselines.bench.benchmarks import *
+from baselines.bench.monitor import *

baselines/bench/benchmarks.py

@@ -1,4 +1,5 @@
 import re
+import os.path as osp
 import os
 SCRIPT_DIR = os.path.dirname(os.path.abspath(__file__))
 

baselines/bench/monitor.py

@@ -1,11 +1,13 @@
 __all__ = ['Monitor', 'get_monitor_files', 'load_results']
 
+import gym
 from gym.core import Wrapper
 import time
 from glob import glob
 import csv
 import os.path as osp
 import json
+import numpy as np
 
 class Monitor(Wrapper):
     EXT = "monitor.csv"
@@ -160,3 +162,27 @@ def load_results(dir):
     df['t'] -= min(header['t_start'] for header in headers)
     df.headers = headers # HACK to preserve backwards compatibility
     return df
+
+def test_monitor():
+    env = gym.make("CartPole-v1")
+    env.seed(0)
+    mon_file = "/tmp/baselines-test-%s.monitor.csv" % uuid.uuid4()
+    menv = Monitor(env, mon_file)
+    menv.reset()
+    for _ in range(1000):
+        _, _, done, _ = menv.step(0)
+        if done:
+            menv.reset()
+
+    f = open(mon_file, 'rt')
+
+    firstline = f.readline()
+    assert firstline.startswith('#')
+    metadata = json.loads(firstline[1:])
+    assert metadata['env_id'] == "CartPole-v1"
+    assert set(metadata.keys()) == {'env_id', 'gym_version', 't_start'},  "Incorrect keys in monitor metadata"
+
+    last_logline = pandas.read_csv(f, index_col=None)
+    assert set(last_logline.keys()) == {'l', 't', 'r'}, "Incorrect keys in monitor logline"
+    f.close()
+    os.remove(mon_file)

baselines/common/atari_wrappers.py

@@ -254,13 +254,6 @@ class LazyFrames(object):
         return len(self._force())
 
     def __getitem__(self, i):
-        return self._force()[i]
-
-    def count(self):
-        frames = self._force()
-        return frames.shape[frames.ndim - 1]
-
-    def frame(self, i):
         return self._force()[..., i]
 
 def make_atari(env_id, max_episode_steps=None):

baselines/common/cmd_util.py

@@ -17,21 +17,26 @@ from baselines.common.atari_wrappers import make_atari, wrap_deepmind
 from baselines.common.vec_env.subproc_vec_env import SubprocVecEnv
 from baselines.common.vec_env.dummy_vec_env import DummyVecEnv
 from baselines.common import retro_wrappers
+from baselines.common.wrappers import ClipActionsWrapper
 
 def make_vec_env(env_id, env_type, num_env, seed,
                  wrapper_kwargs=None,
+                 env_kwargs=None,
                  start_index=0,
                  reward_scale=1.0,
                  flatten_dict_observations=True,
-                 gamestate=None):
+                 gamestate=None,
+                 initializer=None,
+                 force_dummy=False):
     """
     Create a wrapped, monitored SubprocVecEnv for Atari and MuJoCo.
     """
     wrapper_kwargs = wrapper_kwargs or {}
+    env_kwargs = env_kwargs or {}
     mpi_rank = MPI.COMM_WORLD.Get_rank() if MPI else 0
     seed = seed + 10000 * mpi_rank if seed is not None else None
     logger_dir = logger.get_dir()
-    def make_thunk(rank):
+    def make_thunk(rank, initializer=None):
         return lambda: make_env(
             env_id=env_id,
             env_type=env_type,
@@ -42,18 +47,30 @@ def make_vec_env(env_id, env_type, num_env, seed,
             gamestate=gamestate,
             flatten_dict_observations=flatten_dict_observations,
             wrapper_kwargs=wrapper_kwargs,
-            logger_dir=logger_dir
+            env_kwargs=env_kwargs,
+            logger_dir=logger_dir,
+            initializer=initializer
         )
 
     set_global_seeds(seed)
-    if num_env > 1:
-        return SubprocVecEnv([make_thunk(i + start_index) for i in range(num_env)])
+    if not force_dummy and num_env > 1:
+        return SubprocVecEnv([make_thunk(i + start_index, initializer=initializer) for i in range(num_env)])
     else:
-        return DummyVecEnv([make_thunk(start_index)])
+        return DummyVecEnv([make_thunk(i + start_index, initializer=None) for i in range(num_env)])
 
 
-def make_env(env_id, env_type, mpi_rank=0, subrank=0, seed=None, reward_scale=1.0, gamestate=None, flatten_dict_observations=True, wrapper_kwargs=None, logger_dir=None):
+def make_env(env_id, env_type, mpi_rank=0, subrank=0, seed=None, reward_scale=1.0, gamestate=None, flatten_dict_observations=True, wrapper_kwargs=None, env_kwargs=None, logger_dir=None, initializer=None):
+    if initializer is not None:
+        initializer(mpi_rank=mpi_rank, subrank=subrank)
+
     wrapper_kwargs = wrapper_kwargs or {}
+    env_kwargs = env_kwargs or {}
+    if ':' in env_id:
+        import re
+        import importlib
+        module_name = re.sub(':.*','',env_id)
+        env_id = re.sub('.*:', '', env_id)
+        importlib.import_module(module_name)
     if env_type == 'atari':
         env = make_atari(env_id)
     elif env_type == 'retro':
@@ -61,7 +78,7 @@ def make_env(env_id, env_type, mpi_rank=0, subrank=0, seed=None, reward_scale=1.
         gamestate = gamestate or retro.State.DEFAULT
         env = retro_wrappers.make_retro(game=env_id, max_episode_steps=10000, use_restricted_actions=retro.Actions.DISCRETE, state=gamestate)
     else:
-        env = gym.make(env_id)
+        env = gym.make(env_id, **env_kwargs)
 
     if flatten_dict_observations and isinstance(env.observation_space, gym.spaces.Dict):
         keys = env.observation_space.spaces.keys()
@@ -72,6 +89,7 @@ def make_env(env_id, env_type, mpi_rank=0, subrank=0, seed=None, reward_scale=1.
                   logger_dir and os.path.join(logger_dir, str(mpi_rank) + '.' + str(subrank)),
                   allow_early_resets=True)
 
+
     if env_type == 'atari':
         env = wrap_deepmind(env, **wrapper_kwargs)
     elif env_type == 'retro':
@@ -79,6 +97,9 @@ def make_env(env_id, env_type, mpi_rank=0, subrank=0, seed=None, reward_scale=1.
             wrapper_kwargs['frame_stack'] = 1
         env = retro_wrappers.wrap_deepmind_retro(env, **wrapper_kwargs)
 
+    if isinstance(env.action_space, gym.spaces.Box):
+        env = ClipActionsWrapper(env)
+
     if reward_scale != 1:
         env = retro_wrappers.RewardScaler(env, reward_scale)
 
@@ -149,7 +170,6 @@ def common_arg_parser():
     parser.add_argument('--save_path', help='Path to save trained model to', default=None, type=str)
     parser.add_argument('--save_video_interval', help='Save video every x steps (0 = disabled)', default=0, type=int)
     parser.add_argument('--save_video_length', help='Length of recorded video. Default: 200', default=200, type=int)
-    parser.add_argument('--log_path', help='Directory to save learning curve data.', default=None, type=str)
     parser.add_argument('--play', default=False, action='store_true')
     return parser
 
@@ -166,7 +186,7 @@ def robotics_arg_parser():
 
 def parse_unknown_args(args):
     """
-    Parse arguments not consumed by arg parser into a dictionary
+    Parse arguments not consumed by arg parser into a dicitonary
     """
     retval = {}
     preceded_by_key = False

baselines/common/distributions.py

@@ -1,6 +1,8 @@
 import tensorflow as tf
 import numpy as np
+import baselines.common.tf_util as U
 from baselines.a2c.utils import fc
+from tensorflow.python.ops import math_ops
 
 class Pd(object):
     """
@@ -29,7 +31,7 @@ class Pd(object):
     def __getitem__(self, idx):
         return self.__class__(self.flatparam()[idx])
 
-class PdType(tf.Module):
+class PdType(object):
     """
     Parametrized family of probability distributions
     """
@@ -37,7 +39,7 @@ class PdType(tf.Module):
         raise NotImplementedError
     def pdfromflat(self, flat):
         return self.pdclass()(flat)
-    def pdfromlatent(self, latent_vector):
+    def pdfromlatent(self, latent_vector, init_scale, init_bias):
         raise NotImplementedError
     def param_shape(self):
         raise NotImplementedError
@@ -46,18 +48,21 @@ class PdType(tf.Module):
     def sample_dtype(self):
         raise NotImplementedError
 
+    def param_placeholder(self, prepend_shape, name=None):
+        return tf.placeholder(dtype=tf.float32, shape=prepend_shape+self.param_shape(), name=name)
+    def sample_placeholder(self, prepend_shape, name=None):
+        return tf.placeholder(dtype=self.sample_dtype(), shape=prepend_shape+self.sample_shape(), name=name)
+
     def __eq__(self, other):
         return (type(self) == type(other)) and (self.__dict__ == other.__dict__)
 
 class CategoricalPdType(PdType):
-    def __init__(self, latent_shape, ncat, init_scale=1.0, init_bias=0.0):
+    def __init__(self, ncat):
         self.ncat = ncat
-        self.matching_fc = _matching_fc(latent_shape, 'pi', self.ncat, init_scale=init_scale, init_bias=init_bias)
-
     def pdclass(self):
         return CategoricalPd
-    def pdfromlatent(self, latent_vector):
-        pdparam = self.matching_fc(latent_vector)
+    def pdfromlatent(self, latent_vector, init_scale=1.0, init_bias=0.0):
+        pdparam = _matching_fc(latent_vector, 'pi', self.ncat, init_scale=init_scale, init_bias=init_bias)
         return self.pdfromflat(pdparam), pdparam
 
     def param_shape(self):
@@ -67,18 +72,37 @@ class CategoricalPdType(PdType):
     def sample_dtype(self):
         return tf.int32
 
-class DiagGaussianPdType(PdType):
-    def __init__(self, latent_shape, size, init_scale=1.0, init_bias=0.0):
-        self.size = size
-        self.matching_fc = _matching_fc(latent_shape, 'pi', self.size, init_scale=init_scale, init_bias=init_bias)
-        self.logstd = tf.Variable(np.zeros((1, self.size)), name='pi/logstd', dtype=tf.float32)
 
+class MultiCategoricalPdType(PdType):
+    def __init__(self, nvec):
+        self.ncats = nvec.astype('int32')
+        assert (self.ncats > 0).all()
+    def pdclass(self):
+        return MultiCategoricalPd
+    def pdfromflat(self, flat):
+        return MultiCategoricalPd(self.ncats, flat)
+
+    def pdfromlatent(self, latent, init_scale=1.0, init_bias=0.0):
+        pdparam = _matching_fc(latent, 'pi', self.ncats.sum(), init_scale=init_scale, init_bias=init_bias)
+        return self.pdfromflat(pdparam), pdparam
+
+    def param_shape(self):
+        return [sum(self.ncats)]
+    def sample_shape(self):
+        return [len(self.ncats)]
+    def sample_dtype(self):
+        return tf.int32
+
+class DiagGaussianPdType(PdType):
+    def __init__(self, size):
+        self.size = size
     def pdclass(self):
         return DiagGaussianPd
 
-    def pdfromlatent(self, latent_vector):
-        mean = self.matching_fc(latent_vector)
-        pdparam = tf.concat([mean, mean * 0.0 + self.logstd], axis=1)
+    def pdfromlatent(self, latent_vector, init_scale=1.0, init_bias=0.0):
+        mean = _matching_fc(latent_vector, 'pi', self.size, init_scale=init_scale, init_bias=init_bias)
+        logstd = tf.get_variable(name='pi/logstd', shape=[1, self.size], initializer=tf.zeros_initializer())
+        pdparam = tf.concat([mean, mean * 0.0 + logstd], axis=1)
         return self.pdfromflat(pdparam), mean
 
     def param_shape(self):
@@ -88,6 +112,43 @@ class DiagGaussianPdType(PdType):
     def sample_dtype(self):
         return tf.float32
 
+class BernoulliPdType(PdType):
+    def __init__(self, size):
+        self.size = size
+    def pdclass(self):
+        return BernoulliPd
+    def param_shape(self):
+        return [self.size]
+    def sample_shape(self):
+        return [self.size]
+    def sample_dtype(self):
+        return tf.int32
+    def pdfromlatent(self, latent_vector, init_scale=1.0, init_bias=0.0):
+        pdparam = _matching_fc(latent_vector, 'pi', self.size, init_scale=init_scale, init_bias=init_bias)
+        return self.pdfromflat(pdparam), pdparam
+
+# WRONG SECOND DERIVATIVES
+# class CategoricalPd(Pd):
+#     def __init__(self, logits):
+#         self.logits = logits
+#         self.ps = tf.nn.softmax(logits)
+#     @classmethod
+#     def fromflat(cls, flat):
+#         return cls(flat)
+#     def flatparam(self):
+#         return self.logits
+#     def mode(self):
+#         return U.argmax(self.logits, axis=-1)
+#     def logp(self, x):
+#         return -tf.nn.sparse_softmax_cross_entropy_with_logits(self.logits, x)
+#     def kl(self, other):
+#         return tf.nn.softmax_cross_entropy_with_logits(other.logits, self.ps) \
+#                 - tf.nn.softmax_cross_entropy_with_logits(self.logits, self.ps)
+#     def entropy(self):
+#         return tf.nn.softmax_cross_entropy_with_logits(self.logits, self.ps)
+#     def sample(self):
+#         u = tf.random_uniform(tf.shape(self.logits))
+#         return U.argmax(self.logits - tf.log(-tf.log(u)), axis=-1)
 
 class CategoricalPd(Pd):
     def __init__(self, logits):
@@ -100,7 +161,6 @@ class CategoricalPd(Pd):
     @property
     def mean(self):
         return tf.nn.softmax(self.logits)
-
     def neglogp(self, x):
         # return tf.nn.sparse_softmax_cross_entropy_with_logits(logits=self.logits, labels=x)
         # Note: we can't use sparse_softmax_cross_entropy_with_logits because
@@ -116,11 +176,11 @@ class CategoricalPd(Pd):
             x = tf.one_hot(x, self.logits.get_shape().as_list()[-1])
         else:
             # already encoded
-            print('logits is {}'.format(self.logits))
             assert x.shape.as_list() == self.logits.shape.as_list()
 
-        return tf.nn.softmax_cross_entropy_with_logits(logits=self.logits, labels=x)
-
+        return tf.nn.softmax_cross_entropy_with_logits_v2(
+            logits=self.logits,
+            labels=x)
     def kl(self, other):
         a0 = self.logits - tf.reduce_max(self.logits, axis=-1, keepdims=True)
         a1 = other.logits - tf.reduce_max(other.logits, axis=-1, keepdims=True)
@@ -129,20 +189,41 @@ class CategoricalPd(Pd):
         z0 = tf.reduce_sum(ea0, axis=-1, keepdims=True)
         z1 = tf.reduce_sum(ea1, axis=-1, keepdims=True)
         p0 = ea0 / z0
-        return tf.reduce_sum(p0 * (a0 - tf.math.log(z0) - a1 + tf.math.log(z1)), axis=-1)
+        return tf.reduce_sum(p0 * (a0 - tf.log(z0) - a1 + tf.log(z1)), axis=-1)
     def entropy(self):
         a0 = self.logits - tf.reduce_max(self.logits, axis=-1, keepdims=True)
         ea0 = tf.exp(a0)
         z0 = tf.reduce_sum(ea0, axis=-1, keepdims=True)
         p0 = ea0 / z0
-        return tf.reduce_sum(p0 * (tf.math.log(z0) - a0), axis=-1)
+        return tf.reduce_sum(p0 * (tf.log(z0) - a0), axis=-1)
     def sample(self):
-        u = tf.random.uniform(tf.shape(self.logits), dtype=self.logits.dtype, seed=0)
-        return tf.argmax(self.logits - tf.math.log(-tf.math.log(u)), axis=-1)
+        u = tf.random_uniform(tf.shape(self.logits), dtype=self.logits.dtype)
+        return tf.argmax(self.logits - tf.log(-tf.log(u)), axis=-1)
     @classmethod
     def fromflat(cls, flat):
         return cls(flat)
 
+class MultiCategoricalPd(Pd):
+    def __init__(self, nvec, flat):
+        self.flat = flat
+        self.categoricals = list(map(CategoricalPd,
+            tf.split(flat, np.array(nvec, dtype=np.int32), axis=-1)))
+    def flatparam(self):
+        return self.flat
+    def mode(self):
+        return tf.cast(tf.stack([p.mode() for p in self.categoricals], axis=-1), tf.int32)
+    def neglogp(self, x):
+        return tf.add_n([p.neglogp(px) for p, px in zip(self.categoricals, tf.unstack(x, axis=-1))])
+    def kl(self, other):
+        return tf.add_n([p.kl(q) for p, q in zip(self.categoricals, other.categoricals)])
+    def entropy(self):
+        return tf.add_n([p.entropy() for p in self.categoricals])
+    def sample(self):
+        return tf.cast(tf.stack([p.sample() for p in self.categoricals], axis=-1), tf.int32)
+    @classmethod
+    def fromflat(cls, flat):
+        raise NotImplementedError
+
 class DiagGaussianPd(Pd):
     def __init__(self, flat):
         self.flat = flat
@@ -156,7 +237,7 @@ class DiagGaussianPd(Pd):
         return self.mean
     def neglogp(self, x):
         return 0.5 * tf.reduce_sum(tf.square((x - self.mean) / self.std), axis=-1) \
-               + 0.5 * np.log(2.0 * np.pi) * tf.cast(tf.shape(x)[-1], dtype=tf.float32) \
+               + 0.5 * np.log(2.0 * np.pi) * tf.to_float(tf.shape(x)[-1]) \
                + tf.reduce_sum(self.logstd, axis=-1)
     def kl(self, other):
         assert isinstance(other, DiagGaussianPd)
@@ -164,23 +245,111 @@ class DiagGaussianPd(Pd):
     def entropy(self):
         return tf.reduce_sum(self.logstd + .5 * np.log(2.0 * np.pi * np.e), axis=-1)
     def sample(self):
-        return self.mean + self.std * tf.random.normal(tf.shape(self.mean))
+        return self.mean + self.std * tf.random_normal(tf.shape(self.mean))
     @classmethod
     def fromflat(cls, flat):
         return cls(flat)
 
-def make_pdtype(latent_shape, ac_space, init_scale=1.0):
+
+class BernoulliPd(Pd):
+    def __init__(self, logits):
+        self.logits = logits
+        self.ps = tf.sigmoid(logits)
+    def flatparam(self):
+        return self.logits
+    @property
+    def mean(self):
+        return self.ps
+    def mode(self):
+        return tf.round(self.ps)
+    def neglogp(self, x):
+        return tf.reduce_sum(tf.nn.sigmoid_cross_entropy_with_logits(logits=self.logits, labels=tf.to_float(x)), axis=-1)
+    def kl(self, other):
+        return tf.reduce_sum(tf.nn.sigmoid_cross_entropy_with_logits(logits=other.logits, labels=self.ps), axis=-1) - tf.reduce_sum(tf.nn.sigmoid_cross_entropy_with_logits(logits=self.logits, labels=self.ps), axis=-1)
+    def entropy(self):
+        return tf.reduce_sum(tf.nn.sigmoid_cross_entropy_with_logits(logits=self.logits, labels=self.ps), axis=-1)
+    def sample(self):
+        u = tf.random_uniform(tf.shape(self.ps))
+        return tf.to_float(math_ops.less(u, self.ps))
+    @classmethod
+    def fromflat(cls, flat):
+        return cls(flat)
+
+def make_pdtype(ac_space):
     from gym import spaces
     if isinstance(ac_space, spaces.Box):
         assert len(ac_space.shape) == 1
-        return DiagGaussianPdType(latent_shape, ac_space.shape[0], init_scale)
+        return DiagGaussianPdType(ac_space.shape[0])
     elif isinstance(ac_space, spaces.Discrete):
-        return CategoricalPdType(latent_shape, ac_space.n, init_scale)
+        return CategoricalPdType(ac_space.n)
+    elif isinstance(ac_space, spaces.MultiDiscrete):
+        return MultiCategoricalPdType(ac_space.nvec)
+    elif isinstance(ac_space, spaces.MultiBinary):
+        return BernoulliPdType(ac_space.n)
     else:
-        raise ValueError('No implementation for {}'.format(ac_space))
+        raise NotImplementedError
 
-def _matching_fc(tensor_shape, name, size, init_scale, init_bias):
-    if tensor_shape[-1] == size:
-        return lambda x: x
+def shape_el(v, i):
+    maybe = v.get_shape()[i]
+    if maybe is not None:
+        return maybe
     else:
-        return fc(tensor_shape, name, size, init_scale=init_scale, init_bias=init_bias)
+        return tf.shape(v)[i]
+
+@U.in_session
+def test_probtypes():
+    np.random.seed(0)
+
+    pdparam_diag_gauss = np.array([-.2, .3, .4, -.5, .1, -.5, .1, 0.8])
+    diag_gauss = DiagGaussianPdType(pdparam_diag_gauss.size // 2) #pylint: disable=E1101
+    validate_probtype(diag_gauss, pdparam_diag_gauss)
+
+    pdparam_categorical = np.array([-.2, .3, .5])
+    categorical = CategoricalPdType(pdparam_categorical.size) #pylint: disable=E1101
+    validate_probtype(categorical, pdparam_categorical)
+
+    nvec = [1,2,3]
+    pdparam_multicategorical = np.array([-.2, .3, .5, .1, 1, -.1])
+    multicategorical = MultiCategoricalPdType(nvec) #pylint: disable=E1101
+    validate_probtype(multicategorical, pdparam_multicategorical)
+
+    pdparam_bernoulli = np.array([-.2, .3, .5])
+    bernoulli = BernoulliPdType(pdparam_bernoulli.size) #pylint: disable=E1101
+    validate_probtype(bernoulli, pdparam_bernoulli)
+
+
+def validate_probtype(probtype, pdparam):
+    N = 100000
+    # Check to see if mean negative log likelihood == differential entropy
+    Mval = np.repeat(pdparam[None, :], N, axis=0)
+    M = probtype.param_placeholder([N])
+    X = probtype.sample_placeholder([N])
+    pd = probtype.pdfromflat(M)
+    calcloglik = U.function([X, M], pd.logp(X))
+    calcent = U.function([M], pd.entropy())
+    Xval = tf.get_default_session().run(pd.sample(), feed_dict={M:Mval})
+    logliks = calcloglik(Xval, Mval)
+    entval_ll = - logliks.mean() #pylint: disable=E1101
+    entval_ll_stderr = logliks.std() / np.sqrt(N) #pylint: disable=E1101
+    entval = calcent(Mval).mean() #pylint: disable=E1101
+    assert np.abs(entval - entval_ll) < 3 * entval_ll_stderr # within 3 sigmas
+
+    # Check to see if kldiv[p,q] = - ent[p] - E_p[log q]
+    M2 = probtype.param_placeholder([N])
+    pd2 = probtype.pdfromflat(M2)
+    q = pdparam + np.random.randn(pdparam.size) * 0.1
+    Mval2 = np.repeat(q[None, :], N, axis=0)
+    calckl = U.function([M, M2], pd.kl(pd2))
+    klval = calckl(Mval, Mval2).mean() #pylint: disable=E1101
+    logliks = calcloglik(Xval, Mval2)
+    klval_ll = - entval - logliks.mean() #pylint: disable=E1101
+    klval_ll_stderr = logliks.std() / np.sqrt(N) #pylint: disable=E1101
+    assert np.abs(klval - klval_ll) < 3 * klval_ll_stderr # within 3 sigmas
+    print('ok on', probtype, pdparam)
+
+
+def _matching_fc(tensor, name, size, init_scale, init_bias):
+    if tensor.shape[-1] == size:
+        return tensor
+    else:
+        return fc(tensor, name, size, init_scale=init_scale, init_bias=init_bias)

baselines/common/input.py

@@ -0,0 +1,64 @@
+import numpy as np
+import tensorflow as tf
+from gym.spaces import Discrete, Box, MultiDiscrete
+
+def observation_placeholder(ob_space, batch_size=None, name='Ob'):
+    '''
+    Create placeholder to feed observations into of the size appropriate to the observation space
+
+    Parameters:
+    ----------
+
+    ob_space: gym.Space     observation space
+
+    batch_size: int         size of the batch to be fed into input. Can be left None in most cases.
+
+    name: str               name of the placeholder
+
+    Returns:
+    -------
+
+    tensorflow placeholder tensor
+    '''
+
+    assert isinstance(ob_space, Discrete) or isinstance(ob_space, Box) or isinstance(ob_space, MultiDiscrete), \
+        'Can only deal with Discrete and Box observation spaces for now'
+
+    dtype = ob_space.dtype
+    if dtype == np.int8:
+        dtype = np.uint8
+
+    return tf.placeholder(shape=(batch_size,) + ob_space.shape, dtype=dtype, name=name)
+
+
+def observation_input(ob_space, batch_size=None, name='Ob'):
+    '''
+    Create placeholder to feed observations into of the size appropriate to the observation space, and add input
+    encoder of the appropriate type.
+    '''
+
+    placeholder = observation_placeholder(ob_space, batch_size, name)
+    return placeholder, encode_observation(ob_space, placeholder)
+
+def encode_observation(ob_space, placeholder):
+    '''
+    Encode input in the way that is appropriate to the observation space
+
+    Parameters:
+    ----------
+
+    ob_space: gym.Space             observation space
+
+    placeholder: tf.placeholder     observation input placeholder
+    '''
+    if isinstance(ob_space, Discrete):
+        return tf.to_float(tf.one_hot(placeholder, ob_space.n))
+    elif isinstance(ob_space, Box):
+        return tf.to_float(placeholder)
+    elif isinstance(ob_space, MultiDiscrete):
+        placeholder = tf.cast(placeholder, tf.int32)
+        one_hots = [tf.to_float(tf.one_hot(placeholder[..., i], ob_space.nvec[i])) for i in range(placeholder.shape[-1])]
+        return tf.concat(one_hots, axis=-1)
+    else:
+        raise NotImplementedError
+

baselines/common/misc_util.py

@@ -55,7 +55,7 @@ def set_global_seeds(i):
     myseed = i  + 1000 * rank if i is not None else None
     try:
         import tensorflow as tf
-        tf.random.set_seed(myseed)
+        tf.set_random_seed(myseed)
     except ImportError:
         pass
     np.random.seed(myseed)

baselines/common/models.py

@@ -1,6 +1,8 @@
 import numpy as np
 import tensorflow as tf
-from baselines.a2c.utils import ortho_init, conv
+from baselines.a2c import utils
+from baselines.a2c.utils import conv, fc, conv_to_fc, batch_to_seq, seq_to_batch
+from baselines.common.mpi_running_mean_std import RunningMeanStd
 
 mapping = {}
 
@@ -10,26 +12,67 @@ def register(name):
         return func
     return _thunk
 
-
-def nature_cnn(input_shape, **conv_kwargs):
+def nature_cnn(unscaled_images, **conv_kwargs):
     """
     CNN from Nature paper.
     """
-    print('input shape is {}'.format(input_shape))
-    x_input = tf.keras.Input(shape=input_shape, dtype=tf.uint8)
-    h = x_input
-    h = tf.cast(h, tf.float32) / 255.
-    h = conv('c1', nf=32, rf=8, stride=4, activation='relu', init_scale=np.sqrt(2))(h)
-    h2 = conv('c2', nf=64, rf=4, stride=2, activation='relu', init_scale=np.sqrt(2))(h)
-    h3 = conv('c3', nf=64, rf=3, stride=1, activation='relu', init_scale=np.sqrt(2))(h2)
-    h3 = tf.keras.layers.Flatten()(h3)
-    h3 = tf.keras.layers.Dense(units=512, kernel_initializer=ortho_init(np.sqrt(2)),
-                               name='fc1', activation='relu')(h3)
-    network = tf.keras.Model(inputs=[x_input], outputs=[h3])
-    return network
+    scaled_images = tf.cast(unscaled_images, tf.float32) / 255.
+    activ = tf.nn.relu
+    h = activ(conv(scaled_images, 'c1', nf=32, rf=8, stride=4, init_scale=np.sqrt(2),
+                   **conv_kwargs))
+    h2 = activ(conv(h, 'c2', nf=64, rf=4, stride=2, init_scale=np.sqrt(2), **conv_kwargs))
+    h3 = activ(conv(h2, 'c3', nf=64, rf=3, stride=1, init_scale=np.sqrt(2), **conv_kwargs))
+    h3 = conv_to_fc(h3)
+    return activ(fc(h3, 'fc1', nh=512, init_scale=np.sqrt(2)))
+
+def build_impala_cnn(unscaled_images, depths=[16,32,32], **conv_kwargs):
+    """
+    Model used in the paper "IMPALA: Scalable Distributed Deep-RL with
+    Importance Weighted Actor-Learner Architectures" https://arxiv.org/abs/1802.01561
+    """
+
+    layer_num = 0
+
+    def get_layer_num_str():
+        nonlocal layer_num
+        num_str = str(layer_num)
+        layer_num += 1
+        return num_str
+
+    def conv_layer(out, depth):
+        return tf.layers.conv2d(out, depth, 3, padding='same', name='layer_' + get_layer_num_str())
+
+    def residual_block(inputs):
+        depth = inputs.get_shape()[-1].value
+
+        out = tf.nn.relu(inputs)
+
+        out = conv_layer(out, depth)
+        out = tf.nn.relu(out)
+        out = conv_layer(out, depth)
+        return out + inputs
+
+    def conv_sequence(inputs, depth):
+        out = conv_layer(inputs, depth)
+        out = tf.layers.max_pooling2d(out, pool_size=3, strides=2, padding='same')
+        out = residual_block(out)
+        out = residual_block(out)
+        return out
+
+    out = tf.cast(unscaled_images, tf.float32) / 255.
+
+    for depth in depths:
+        out = conv_sequence(out, depth)
+
+    out = tf.layers.flatten(out)
+    out = tf.nn.relu(out)
+    out = tf.layers.dense(out, 256, activation=tf.nn.relu, name='layer_' + get_layer_num_str())
+
+    return out
+
 
 @register("mlp")
-def mlp(num_layers=2, num_hidden=64, activation=tf.tanh):
+def mlp(num_layers=2, num_hidden=64, activation=tf.tanh, layer_norm=False):
     """
     Stack of fully-connected layers to be used in a policy / q-function approximator
 
@@ -47,54 +90,169 @@ def mlp(num_layers=2, num_hidden=64, activation=tf.tanh):
 
     function that builds fully connected network with a given input tensor / placeholder
     """
-    def network_fn(input_shape):
-        print('input shape is {}'.format(input_shape))
-        x_input = tf.keras.Input(shape=input_shape)
-        # h = tf.keras.layers.Flatten(x_input)
-        h = x_input
+    def network_fn(X):
+        h = tf.layers.flatten(X)
         for i in range(num_layers):
-          h = tf.keras.layers.Dense(units=num_hidden, kernel_initializer=ortho_init(np.sqrt(2)),
-                                    name='mlp_fc{}'.format(i), activation=activation)(h)
+            h = fc(h, 'mlp_fc{}'.format(i), nh=num_hidden, init_scale=np.sqrt(2))
+            if layer_norm:
+                h = tf.contrib.layers.layer_norm(h, center=True, scale=True)
+            h = activation(h)
 
-        network = tf.keras.Model(inputs=[x_input], outputs=[h])
-        return network
+        return h
 
     return network_fn
 
 
 @register("cnn")
 def cnn(**conv_kwargs):
-    def network_fn(input_shape):
-        return nature_cnn(input_shape, **conv_kwargs)
+    def network_fn(X):
+        return nature_cnn(X, **conv_kwargs)
     return network_fn
 
+@register("impala_cnn")
+def impala_cnn(**conv_kwargs):
+    def network_fn(X):
+        return build_impala_cnn(X)
+    return network_fn
+
+@register("cnn_small")
+def cnn_small(**conv_kwargs):
+    def network_fn(X):
+        h = tf.cast(X, tf.float32) / 255.
+
+        activ = tf.nn.relu
+        h = activ(conv(h, 'c1', nf=8, rf=8, stride=4, init_scale=np.sqrt(2), **conv_kwargs))
+        h = activ(conv(h, 'c2', nf=16, rf=4, stride=2, init_scale=np.sqrt(2), **conv_kwargs))
+        h = conv_to_fc(h)
+        h = activ(fc(h, 'fc1', nh=128, init_scale=np.sqrt(2)))
+        return h
+    return network_fn
+
+@register("lstm")
+def lstm(nlstm=128, layer_norm=False):
+    """
+    Builds LSTM (Long-Short Term Memory) network to be used in a policy.
+    Note that the resulting function returns not only the output of the LSTM
+    (i.e. hidden state of lstm for each step in the sequence), but also a dictionary
+    with auxiliary tensors to be set as policy attributes.
+
+    Specifically,
+        S is a placeholder to feed current state (LSTM state has to be managed outside policy)
+        M is a placeholder for the mask (used to mask out observations after the end of the episode, but can be used for other purposes too)
+        initial_state is a numpy array containing initial lstm state (usually zeros)
+        state is the output LSTM state (to be fed into S at the next call)
+
+
+    An example of usage of lstm-based policy can be found here: common/tests/test_doc_examples.py/test_lstm_example
+
+    Parameters:
+    ----------
+
+    nlstm: int          LSTM hidden state size
+
+    layer_norm: bool    if True, layer-normalized version of LSTM is used
+
+    Returns:
+    -------
+
+    function that builds LSTM with a given input tensor / placeholder
+    """
+
+    def network_fn(X, nenv=1):
+        nbatch = X.shape[0]
+        nsteps = nbatch // nenv
+
+        h = tf.layers.flatten(X)
+
+        M = tf.placeholder(tf.float32, [nbatch]) #mask (done t-1)
+        S = tf.placeholder(tf.float32, [nenv, 2*nlstm]) #states
+
+        xs = batch_to_seq(h, nenv, nsteps)
+        ms = batch_to_seq(M, nenv, nsteps)
+
+        if layer_norm:
+            h5, snew = utils.lnlstm(xs, ms, S, scope='lnlstm', nh=nlstm)
+        else:
+            h5, snew = utils.lstm(xs, ms, S, scope='lstm', nh=nlstm)
+
+        h = seq_to_batch(h5)
+        initial_state = np.zeros(S.shape.as_list(), dtype=float)
+
+        return h, {'S':S, 'M':M, 'state':snew, 'initial_state':initial_state}
+
+    return network_fn
+
+
+@register("cnn_lstm")
+def cnn_lstm(nlstm=128, layer_norm=False, conv_fn=nature_cnn, **conv_kwargs):
+    def network_fn(X, nenv=1):
+        nbatch = X.shape[0]
+        nsteps = nbatch // nenv
+
+        h = conv_fn(X, **conv_kwargs)
+
+        M = tf.placeholder(tf.float32, [nbatch]) #mask (done t-1)
+        S = tf.placeholder(tf.float32, [nenv, 2*nlstm]) #states
+
+        xs = batch_to_seq(h, nenv, nsteps)
+        ms = batch_to_seq(M, nenv, nsteps)
+
+        if layer_norm:
+            h5, snew = utils.lnlstm(xs, ms, S, scope='lnlstm', nh=nlstm)
+        else:
+            h5, snew = utils.lstm(xs, ms, S, scope='lstm', nh=nlstm)
+
+        h = seq_to_batch(h5)
+        initial_state = np.zeros(S.shape.as_list(), dtype=float)
+
+        return h, {'S':S, 'M':M, 'state':snew, 'initial_state':initial_state}
+
+    return network_fn
+
+@register("impala_cnn_lstm")
+def impala_cnn_lstm():
+    return cnn_lstm(nlstm=256, conv_fn=build_impala_cnn)
+
+@register("cnn_lnlstm")
+def cnn_lnlstm(nlstm=128, **conv_kwargs):
+    return cnn_lstm(nlstm, layer_norm=True, **conv_kwargs)
+
 
 @register("conv_only")
 def conv_only(convs=[(32, 8, 4), (64, 4, 2), (64, 3, 1)], **conv_kwargs):
     '''
     convolutions-only net
+
     Parameters:
     ----------
+
     conv:       list of triples (filter_number, filter_size, stride) specifying parameters for each layer.
+
     Returns:
+
     function that takes tensorflow tensor as input and returns the output of the last convolutional layer
+
     '''
 
-    def network_fn(input_shape):
-        print('input shape is {}'.format(input_shape))
-        x_input = tf.keras.Input(shape=input_shape, dtype=tf.uint8)
-        h = x_input
-        h = tf.cast(h, tf.float32) / 255.
-        with tf.name_scope("convnet"):
+    def network_fn(X):
+        out = tf.cast(X, tf.float32) / 255.
+        with tf.variable_scope("convnet"):
             for num_outputs, kernel_size, stride in convs:
-                h = tf.keras.layers.Conv2D(
-                    filters=num_outputs, kernel_size=kernel_size, strides=stride,
-                    activation='relu', **conv_kwargs)(h)
+                out = tf.contrib.layers.convolution2d(out,
+                                           num_outputs=num_outputs,
+                                           kernel_size=kernel_size,
+                                           stride=stride,
+                                           activation_fn=tf.nn.relu,
+                                           **conv_kwargs)
 
-        network = tf.keras.Model(inputs=[x_input], outputs=[h])
-        return network
+        return out
     return network_fn
 
+def _normalize_clip_observation(x, clip_range=[-5.0, 5.0]):
+    rms = RunningMeanStd(shape=x.shape[1:])
+    norm_x = tf.clip_by_value((x - rms.mean) / rms.std, min(clip_range), max(clip_range))
+    return norm_x, rms
+
 
 def get_network_builder(name):
     """

baselines/common/mpi_adam.py

@@ -1,4 +1,5 @@
 import baselines.common.tf_util as U
+import tensorflow as tf
 import numpy as np
 try:
     from mpi4py import MPI
@@ -58,3 +59,45 @@ class MpiAdam(object):
             thetaroot = np.empty_like(thetalocal)
             self.comm.Bcast(thetaroot, root=0)
             assert (thetaroot == thetalocal).all(), (thetaroot, thetalocal)
+
+@U.in_session
+def test_MpiAdam():
+    np.random.seed(0)
+    tf.set_random_seed(0)
+
+    a = tf.Variable(np.random.randn(3).astype('float32'))
+    b = tf.Variable(np.random.randn(2,5).astype('float32'))
+    loss = tf.reduce_sum(tf.square(a)) + tf.reduce_sum(tf.sin(b))
+
+    stepsize = 1e-2
+    update_op = tf.train.AdamOptimizer(stepsize).minimize(loss)
+    do_update = U.function([], loss, updates=[update_op])
+
+    tf.get_default_session().run(tf.global_variables_initializer())
+    losslist_ref = []
+    for i in range(10):
+        l = do_update()
+        print(i, l)
+        losslist_ref.append(l)
+
+
+
+    tf.set_random_seed(0)
+    tf.get_default_session().run(tf.global_variables_initializer())
+
+    var_list = [a,b]
+    lossandgrad = U.function([], [loss, U.flatgrad(loss, var_list)])
+    adam = MpiAdam(var_list)
+
+    losslist_test = []
+    for i in range(10):
+        l,g = lossandgrad()
+        adam.update(g, stepsize)
+        print(i,l)
+        losslist_test.append(l)
+
+    np.testing.assert_allclose(np.array(losslist_ref), np.array(losslist_test), atol=1e-4)
+
+
+if __name__ == '__main__':
+    test_MpiAdam()

baselines/common/mpi_adam_optimizer.py

@@ -1,45 +1,54 @@
 import numpy as np
 import tensorflow as tf
+from baselines.common import tf_util as U
+from baselines.common.tests.test_with_mpi import with_mpi
+from baselines import logger
 try:
     from mpi4py import MPI
 except ImportError:
     MPI = None
 
-class MpiAdamOptimizer(tf.Module):
+class MpiAdamOptimizer(tf.train.AdamOptimizer):
     """Adam optimizer that averages gradients across mpi processes."""
-    def __init__(self, comm, var_list):
-        self.var_list = var_list
+    def __init__(self, comm, grad_clip=None, mpi_rank_weight=1, **kwargs):
         self.comm = comm
-        self.beta1 = 0.9
-        self.beta2 = 0.999
-        self.epsilon = 1e-08
-        self.t = tf.Variable(0, name='step', dtype=tf.int32)
-        var_shapes = [v.shape.as_list() for v in var_list]
-        self.var_sizes = [int(np.prod(s)) for s in var_shapes]
-        self.flat_var_size = sum(self.var_sizes)
-        self.m = tf.Variable(np.zeros(self.flat_var_size, 'float32'))
-        self.v = tf.Variable(np.zeros(self.flat_var_size, 'float32'))
+        self.grad_clip = grad_clip
+        self.mpi_rank_weight = mpi_rank_weight
+        tf.train.AdamOptimizer.__init__(self, **kwargs)
+    def compute_gradients(self, loss, var_list, **kwargs):
+        grads_and_vars = tf.train.AdamOptimizer.compute_gradients(self, loss, var_list, **kwargs)
+        grads_and_vars = [(g, v) for g, v in grads_and_vars if g is not None]
+        flat_grad = tf.concat([tf.reshape(g, (-1,)) for g, v in grads_and_vars], axis=0) * self.mpi_rank_weight
+        shapes = [v.shape.as_list() for g, v in grads_and_vars]
+        sizes = [int(np.prod(s)) for s in shapes]
 
-    def apply_gradients(self, flat_grad, lr):
-        buf = np.zeros(self.flat_var_size, np.float32)
-        self.comm.Allreduce(flat_grad.numpy(), buf, op=MPI.SUM)
-        avg_flat_grad = np.divide(buf, float(self.comm.Get_size()))
-        self._apply_gradients(tf.constant(avg_flat_grad), lr)
-        if self.t.numpy() % 100 == 0:
-            check_synced(tf.reduce_sum(self.var_list[0]).numpy())
+        total_weight = np.zeros(1, np.float32)
+        self.comm.Allreduce(np.array([self.mpi_rank_weight], dtype=np.float32), total_weight, op=MPI.SUM)
+        total_weight = total_weight[0]
 
-    @tf.function
-    def _apply_gradients(self, avg_flat_grad, lr):
-        self.t.assign_add(1)
-        t = tf.cast(self.t, tf.float32)
-        a = lr * tf.math.sqrt(1 - tf.math.pow(self.beta2, t)) / (1 - tf.math.pow(self.beta1, t))
-        self.m.assign(self.beta1 * self.m + (1 - self.beta1) * avg_flat_grad)
-        self.v.assign(self.beta2 * self.v + (1 - self.beta2) * tf.math.square(avg_flat_grad))
-        flat_step = (- a) * self.m / (tf.math.sqrt(self.v) + self.epsilon)
-        var_steps = tf.split(flat_step, self.var_sizes, axis=0)
-        for var_step, var in zip(var_steps, self.var_list):
-            var.assign_add(tf.reshape(var_step, var.shape))
+        buf = np.zeros(sum(sizes), np.float32)
+        countholder = [0] # Counts how many times _collect_grads has been called
+        stat = tf.reduce_sum(grads_and_vars[0][1]) # sum of first variable
+        def _collect_grads(flat_grad, np_stat):
+            if self.grad_clip is not None:
+                gradnorm = np.linalg.norm(flat_grad)
+                if gradnorm > 1:
+                    flat_grad /= gradnorm
+                logger.logkv_mean('gradnorm', gradnorm)
+                logger.logkv_mean('gradclipfrac', float(gradnorm > 1))
+            self.comm.Allreduce(flat_grad, buf, op=MPI.SUM)
+            np.divide(buf, float(total_weight), out=buf)
+            if countholder[0] % 100 == 0:
+                check_synced(np_stat, self.comm)
+            countholder[0] += 1
+            return buf
 
+        avg_flat_grad = tf.py_func(_collect_grads, [flat_grad, stat], tf.float32)
+        avg_flat_grad.set_shape(flat_grad.shape)
+        avg_grads = tf.split(avg_flat_grad, sizes, axis=0)
+        avg_grads_and_vars = [(tf.reshape(g, v.shape), v)
+                    for g, (_, v) in zip(avg_grads, grads_and_vars)]
+        return avg_grads_and_vars
 
 def check_synced(localval, comm=None):
     """
@@ -56,4 +65,26 @@ def check_synced(localval, comm=None):
     vals = comm.gather(localval)
     if comm.rank == 0:
         assert all(val==vals[0] for val in vals[1:]),\
-            'MpiAdamOptimizer detected that different workers have different weights: {}'.format(vals)
+            f'MpiAdamOptimizer detected that different workers have different weights: {vals}'
+
+@with_mpi(timeout=5)
+def test_nonfreeze():
+    np.random.seed(0)
+    tf.set_random_seed(0)
+
+    a = tf.Variable(np.random.randn(3).astype('float32'))
+    b = tf.Variable(np.random.randn(2,5).astype('float32'))
+    loss = tf.reduce_sum(tf.square(a)) + tf.reduce_sum(tf.sin(b))
+
+    stepsize = 1e-2
+    # for some reason the session config with inter_op_parallelism_threads was causing
+    # nested sess.run calls to freeze
+    config = tf.ConfigProto(inter_op_parallelism_threads=1)
+    sess = U.get_session(config=config)
+    update_op = MpiAdamOptimizer(comm=MPI.COMM_WORLD, learning_rate=stepsize).minimize(loss)
+    sess.run(tf.global_variables_initializer())
+    losslist_ref = []
+    for i in range(100):
+        l,_ = sess.run([loss, update_op])
+        print(i, l)
+        losslist_ref.append(l)

baselines/common/mpi_running_mean_std.py

@@ -3,54 +3,110 @@ try:
 except ImportError:
     MPI = None
 
-import tensorflow as tf, numpy as np
+import tensorflow as tf, baselines.common.tf_util as U, numpy as np
 
-class RunningMeanStd(tf.Module):
+class RunningMeanStd(object):
     # https://en.wikipedia.org/wiki/Algorithms_for_calculating_variance#Parallel_algorithm
-    def __init__(self, epsilon=1e-2, shape=(), default_clip_range=np.inf):
+    def __init__(self, epsilon=1e-2, shape=()):
 
-        self._sum = tf.Variable(
-            initial_value=np.zeros(shape=shape, dtype=np.float64),
+        self._sum = tf.get_variable(
             dtype=tf.float64,
+            shape=shape,
+            initializer=tf.constant_initializer(0.0),
             name="runningsum", trainable=False)
-        self._sumsq = tf.Variable(
-            initial_value=np.full(shape=shape, fill_value=epsilon, dtype=np.float64),
+        self._sumsq = tf.get_variable(
             dtype=tf.float64,
+            shape=shape,
+            initializer=tf.constant_initializer(epsilon),
             name="runningsumsq", trainable=False)
-        self._count = tf.Variable(
-            initial_value=epsilon,
+        self._count = tf.get_variable(
             dtype=tf.float64,
+            shape=(),
+            initializer=tf.constant_initializer(epsilon),
             name="count", trainable=False)
         self.shape = shape
-        self.epsilon = epsilon
-        self.default_clip_range = default_clip_range
+
+        self.mean = tf.to_float(self._sum / self._count)
+        self.std = tf.sqrt( tf.maximum( tf.to_float(self._sumsq / self._count) - tf.square(self.mean) , 1e-2 ))
+
+        newsum = tf.placeholder(shape=self.shape, dtype=tf.float64, name='sum')
+        newsumsq = tf.placeholder(shape=self.shape, dtype=tf.float64, name='var')
+        newcount = tf.placeholder(shape=[], dtype=tf.float64, name='count')
+        self.incfiltparams = U.function([newsum, newsumsq, newcount], [],
+            updates=[tf.assign_add(self._sum, newsum),
+                     tf.assign_add(self._sumsq, newsumsq),
+                     tf.assign_add(self._count, newcount)])
+
 
     def update(self, x):
         x = x.astype('float64')
         n = int(np.prod(self.shape))
-        addvec = np.concatenate([x.sum(axis=0).ravel(), np.square(x).sum(axis=0).ravel(), np.array([len(x)],dtype='float64')])
         totalvec = np.zeros(n*2+1, 'float64')
+        addvec = np.concatenate([x.sum(axis=0).ravel(), np.square(x).sum(axis=0).ravel(), np.array([len(x)],dtype='float64')])
         if MPI is not None:
-            # totalvec = np.zeros(n*2+1, 'float64')
             MPI.COMM_WORLD.Allreduce(addvec, totalvec, op=MPI.SUM)
-        # else:
-        #     totalvec = addvec
-        self._sum.assign_add(totalvec[0:n].reshape(self.shape))
-        self._sumsq.assign_add(totalvec[n:2*n].reshape(self.shape))
-        self._count.assign_add(totalvec[2*n])
+        self.incfiltparams(totalvec[0:n].reshape(self.shape), totalvec[n:2*n].reshape(self.shape), totalvec[2*n])
 
-    @property
-    def mean(self):
-        return tf.cast(self._sum / self._count, tf.float32)
+@U.in_session
+def test_runningmeanstd():
+    for (x1, x2, x3) in [
+        (np.random.randn(3), np.random.randn(4), np.random.randn(5)),
+        (np.random.randn(3,2), np.random.randn(4,2), np.random.randn(5,2)),
+        ]:
 
-    @property
-    def std(self):
-        return tf.sqrt(tf.maximum(tf.cast(self._sumsq / self._count, tf.float32) - tf.square(self.mean), self.epsilon))
+        rms = RunningMeanStd(epsilon=0.0, shape=x1.shape[1:])
+        U.initialize()
 
-    def normalize(self, v, clip_range=None):
-        if clip_range is None:
-            clip_range = self.default_clip_range
-        return tf.clip_by_value((v - self.mean) / self.std, -clip_range, clip_range)
+        x = np.concatenate([x1, x2, x3], axis=0)
+        ms1 = [x.mean(axis=0), x.std(axis=0)]
+        rms.update(x1)
+        rms.update(x2)
+        rms.update(x3)
+        ms2 = [rms.mean.eval(), rms.std.eval()]
 
-    def denormalize(self, v):
-        return self.mean + v * self.std
+        assert np.allclose(ms1, ms2)
+
+@U.in_session
+def test_dist():
+    np.random.seed(0)
+    p1,p2,p3=(np.random.randn(3,1), np.random.randn(4,1), np.random.randn(5,1))
+    q1,q2,q3=(np.random.randn(6,1), np.random.randn(7,1), np.random.randn(8,1))
+
+    # p1,p2,p3=(np.random.randn(3), np.random.randn(4), np.random.randn(5))
+    # q1,q2,q3=(np.random.randn(6), np.random.randn(7), np.random.randn(8))
+
+    comm = MPI.COMM_WORLD
+    assert comm.Get_size()==2
+    if comm.Get_rank()==0:
+        x1,x2,x3 = p1,p2,p3
+    elif comm.Get_rank()==1:
+        x1,x2,x3 = q1,q2,q3
+    else:
+        assert False
+
+    rms = RunningMeanStd(epsilon=0.0, shape=(1,))
+    U.initialize()
+
+    rms.update(x1)
+    rms.update(x2)
+    rms.update(x3)
+
+    bigvec = np.concatenate([p1,p2,p3,q1,q2,q3])
+
+    def checkallclose(x,y):
+        print(x,y)
+        return np.allclose(x,y)
+
+    assert checkallclose(
+        bigvec.mean(axis=0),
+        rms.mean.eval(),
+    )
+    assert checkallclose(
+        bigvec.std(axis=0),
+        rms.std.eval(),
+    )
+
+
+if __name__ == "__main__":
+    # Run with mpirun -np 2 python <filename>
+    test_dist()

baselines/common/mpi_util.py

@@ -12,16 +12,18 @@ except ImportError:
     MPI = None
 
 
-def sync_from_root(variables, comm=None):
+def sync_from_root(sess, variables, comm=None):
     """
     Send the root node's parameters to every worker.
     Arguments:
+      sess: the TensorFlow session.
       variables: all parameter variables including optimizer's
     """
     if comm is None: comm = MPI.COMM_WORLD
-    values = comm.bcast([var.numpy() for var in variables])
-    for (var, val) in zip(variables, values):
-        var.assign(val)
+    import tensorflow as tf
+    values = comm.bcast(sess.run(variables))
+    sess.run([tf.assign(var, val)
+        for (var, val) in zip(variables, values)])
 
 def gpu_count():
     """

baselines/common/policies.py

@@ -1,68 +1,101 @@
 import tensorflow as tf
+from baselines.common import tf_util
 from baselines.a2c.utils import fc
 from baselines.common.distributions import make_pdtype
+from baselines.common.input import observation_placeholder, encode_observation
+from baselines.common.tf_util import adjust_shape
+from baselines.common.mpi_running_mean_std import RunningMeanStd
+from baselines.common.models import get_network_builder
 
 import gym
 
 
-class PolicyWithValue(tf.Module):
+class PolicyWithValue(object):
     """
     Encapsulates fields and methods for RL policy and value function estimation with shared parameters
     """
 
-    def __init__(self, ac_space, policy_network, value_network=None, estimate_q=False):
+    def __init__(self, env, observations, latent, estimate_q=False, vf_latent=None, sess=None, **tensors):
         """
         Parameters:
         ----------
-        ac_space        action space
+        env             RL environment
 
-        policy_network  keras network for policy
+        observations    tensorflow placeholder in which the observations will be fed
 
-        value_network   keras network for value
+        latent          latent state from which policy distribution parameters should be inferred
 
-        estimate_q      q value or v value
+        vf_latent       latent state from which value function should be inferred (if None, then latent is used)
+
+        sess            tensorflow session to run calculations in (if None, default session is used)
+
+        **tensors       tensorflow tensors for additional attributes such as state or mask
 
         """
 
-        self.policy_network = policy_network
-        self.value_network = value_network or policy_network
-        self.estimate_q = estimate_q
+        self.X = observations
+        self.state = tf.constant([])
         self.initial_state = None
+        self.__dict__.update(tensors)
+
+        vf_latent = vf_latent if vf_latent is not None else latent
+
+        vf_latent = tf.layers.flatten(vf_latent)
+        latent = tf.layers.flatten(latent)
 
         # Based on the action space, will select what probability distribution type
-        self.pdtype = make_pdtype(policy_network.output_shape, ac_space, init_scale=0.01)
+        self.pdtype = make_pdtype(env.action_space)
+
+        self.pd, self.pi = self.pdtype.pdfromlatent(latent, init_scale=0.01)
+
+        # Take an action
+        self.action = self.pd.sample()
+
+        # Calculate the neg log of our probability
+        self.neglogp = self.pd.neglogp(self.action)
+        self.sess = sess or tf.get_default_session()
 
         if estimate_q:
-            assert isinstance(ac_space, gym.spaces.Discrete)
-            self.value_fc = fc(self.value_network.output_shape, 'q', ac_space.n)
+            assert isinstance(env.action_space, gym.spaces.Discrete)
+            self.q = fc(vf_latent, 'q', env.action_space.n)
+            self.vf = self.q
         else:
-            self.value_fc = fc(self.value_network.output_shape, 'vf', 1)
+            self.vf = fc(vf_latent, 'vf', 1)
+            self.vf = self.vf[:,0]
 
-    @tf.function
-    def step(self, observation):
+    def _evaluate(self, variables, observation, **extra_feed):
+        sess = self.sess
+        feed_dict = {self.X: adjust_shape(self.X, observation)}
+        for inpt_name, data in extra_feed.items():
+            if inpt_name in self.__dict__.keys():
+                inpt = self.__dict__[inpt_name]
+                if isinstance(inpt, tf.Tensor) and inpt._op.type == 'Placeholder':
+                    feed_dict[inpt] = adjust_shape(inpt, data)
+
+        return sess.run(variables, feed_dict)
+
+    def step(self, observation, **extra_feed):
         """
         Compute next action(s) given the observation(s)
 
         Parameters:
         ----------
 
-        observation     batched observation data
+        observation     observation data (either single or a batch)
+
+        **extra_feed    additional data such as state or mask (names of the arguments should match the ones in constructor, see __init__)
 
         Returns:
         -------
         (action, value estimate, next state, negative log likelihood of the action under current policy parameters) tuple
         """
 
-        latent = self.policy_network(observation)
-        pd, pi = self.pdtype.pdfromlatent(latent)
-        action = pd.sample()
-        neglogp = pd.neglogp(action)
-        value_latent = self.value_network(observation)
-        vf = tf.squeeze(self.value_fc(value_latent), axis=1)
-        return action, vf, None, neglogp
+        a, v, state, neglogp = self._evaluate([self.action, self.vf, self.state, self.neglogp], observation, **extra_feed)
+        if state.size == 0:
+            state = None
+        return a, v, state, neglogp
 
-    @tf.function
-    def value(self, observation):
+    def value(self, ob, *args, **kwargs):
         """
         Compute value estimate(s) given the observation(s)
 
@@ -71,11 +104,83 @@ class PolicyWithValue(tf.Module):
 
         observation     observation data (either single or a batch)
 
+        **extra_feed    additional data such as state or mask (names of the arguments should match the ones in constructor, see __init__)
+
         Returns:
         -------
         value estimate
         """
-        value_latent = self.value_network(observation)
-        result = tf.squeeze(self.value_fc(value_latent), axis=1)
-        return result
+        return self._evaluate(self.vf, ob, *args, **kwargs)
+
+    def save(self, save_path):
+        tf_util.save_state(save_path, sess=self.sess)
+
+    def load(self, load_path):
+        tf_util.load_state(load_path, sess=self.sess)
+
+def build_policy(env, policy_network, value_network=None,  normalize_observations=False, estimate_q=False, **policy_kwargs):
+    if isinstance(policy_network, str):
+        network_type = policy_network
+        policy_network = get_network_builder(network_type)(**policy_kwargs)
+
+    def policy_fn(nbatch=None, nsteps=None, sess=None, observ_placeholder=None):
+        ob_space = env.observation_space
+
+        X = observ_placeholder if observ_placeholder is not None else observation_placeholder(ob_space, batch_size=nbatch)
+
+        extra_tensors = {}
+
+        if normalize_observations and X.dtype == tf.float32:
+            encoded_x, rms = _normalize_clip_observation(X)
+            extra_tensors['rms'] = rms
+        else:
+            encoded_x = X
+
+        encoded_x = encode_observation(ob_space, encoded_x)
+
+        with tf.variable_scope('pi', reuse=tf.AUTO_REUSE):
+            policy_latent = policy_network(encoded_x)
+            if isinstance(policy_latent, tuple):
+                policy_latent, recurrent_tensors = policy_latent
+
+                if recurrent_tensors is not None:
+                    # recurrent architecture, need a few more steps
+                    nenv = nbatch // nsteps
+                    assert nenv > 0, 'Bad input for recurrent policy: batch size {} smaller than nsteps {}'.format(nbatch, nsteps)
+                    policy_latent, recurrent_tensors = policy_network(encoded_x, nenv)
+                    extra_tensors.update(recurrent_tensors)
+
+
+        _v_net = value_network
+
+        if _v_net is None or _v_net == 'shared':
+            vf_latent = policy_latent
+        else:
+            if _v_net == 'copy':
+                _v_net = policy_network
+            else:
+                assert callable(_v_net)
+
+            with tf.variable_scope('vf', reuse=tf.AUTO_REUSE):
+                # TODO recurrent architectures are not supported with value_network=copy yet
+                vf_latent = _v_net(encoded_x)
+
+        policy = PolicyWithValue(
+            env=env,
+            observations=X,
+            latent=policy_latent,
+            vf_latent=vf_latent,
+            sess=sess,
+            estimate_q=estimate_q,
+            **extra_tensors
+        )
+        return policy
+
+    return policy_fn
+
+
+def _normalize_clip_observation(x, clip_range=[-5.0, 5.0]):
+    rms = RunningMeanStd(shape=x.shape[1:])
+    norm_x = tf.clip_by_value((x - rms.mean) / rms.std, min(clip_range), max(clip_range))
+    return norm_x, rms
 

baselines/common/running_mean_std.py

@@ -1,4 +1,6 @@
+import tensorflow as tf
 import numpy as np
+from baselines.common.tf_util import get_session
 
 class RunningMeanStd(object):
     # https://en.wikipedia.org/wiki/Algorithms_for_calculating_variance#Parallel_algorithm
@@ -29,3 +31,157 @@ def update_mean_var_count_from_moments(mean, var, count, batch_mean, batch_var,
     new_count = tot_count
 
     return new_mean, new_var, new_count
+
+
+class TfRunningMeanStd(object):
+    # https://en.wikipedia.org/wiki/Algorithms_for_calculating_variance#Parallel_algorithm
+    '''
+    TensorFlow variables-based implmentation of computing running mean and std
+    Benefit of this implementation is that it can be saved / loaded together with the tensorflow model
+    '''
+    def __init__(self, epsilon=1e-4, shape=(), scope=''):
+        sess = get_session()
+
+        self._new_mean = tf.placeholder(shape=shape, dtype=tf.float64)
+        self._new_var = tf.placeholder(shape=shape, dtype=tf.float64)
+        self._new_count = tf.placeholder(shape=(), dtype=tf.float64)
+
+
+        with tf.variable_scope(scope, reuse=tf.AUTO_REUSE):
+            self._mean  = tf.get_variable('mean',  initializer=np.zeros(shape, 'float64'),      dtype=tf.float64)
+            self._var   = tf.get_variable('std',   initializer=np.ones(shape, 'float64'),       dtype=tf.float64)
+            self._count = tf.get_variable('count', initializer=np.full((), epsilon, 'float64'), dtype=tf.float64)
+
+        self.update_ops = tf.group([
+            self._var.assign(self._new_var),
+            self._mean.assign(self._new_mean),
+            self._count.assign(self._new_count)
+        ])
+
+        sess.run(tf.variables_initializer([self._mean, self._var, self._count]))
+        self.sess = sess
+        self._set_mean_var_count()
+
+    def _set_mean_var_count(self):
+        self.mean, self.var, self.count = self.sess.run([self._mean, self._var, self._count])
+
+    def update(self, x):
+        batch_mean = np.mean(x, axis=0)
+        batch_var = np.var(x, axis=0)
+        batch_count = x.shape[0]
+
+        new_mean, new_var, new_count = update_mean_var_count_from_moments(self.mean, self.var, self.count, batch_mean, batch_var, batch_count)
+
+        self.sess.run(self.update_ops, feed_dict={
+            self._new_mean: new_mean,
+            self._new_var: new_var,
+            self._new_count: new_count
+        })
+
+        self._set_mean_var_count()
+
+
+
+def test_runningmeanstd():
+    for (x1, x2, x3) in [
+        (np.random.randn(3), np.random.randn(4), np.random.randn(5)),
+        (np.random.randn(3,2), np.random.randn(4,2), np.random.randn(5,2)),
+        ]:
+
+        rms = RunningMeanStd(epsilon=0.0, shape=x1.shape[1:])
+
+        x = np.concatenate([x1, x2, x3], axis=0)
+        ms1 = [x.mean(axis=0), x.var(axis=0)]
+        rms.update(x1)
+        rms.update(x2)
+        rms.update(x3)
+        ms2 = [rms.mean, rms.var]
+
+        np.testing.assert_allclose(ms1, ms2)
+
+def test_tf_runningmeanstd():
+    for (x1, x2, x3) in [
+        (np.random.randn(3), np.random.randn(4), np.random.randn(5)),
+        (np.random.randn(3,2), np.random.randn(4,2), np.random.randn(5,2)),
+        ]:
+
+        rms = TfRunningMeanStd(epsilon=0.0, shape=x1.shape[1:], scope='running_mean_std' + str(np.random.randint(0, 128)))
+
+        x = np.concatenate([x1, x2, x3], axis=0)
+        ms1 = [x.mean(axis=0), x.var(axis=0)]
+        rms.update(x1)
+        rms.update(x2)
+        rms.update(x3)
+        ms2 = [rms.mean, rms.var]
+
+        np.testing.assert_allclose(ms1, ms2)
+
+
+def profile_tf_runningmeanstd():
+    import time
+    from baselines.common import tf_util
+
+    tf_util.get_session( config=tf.ConfigProto(
+        inter_op_parallelism_threads=1,
+        intra_op_parallelism_threads=1,
+        allow_soft_placement=True
+    ))
+
+    x = np.random.random((376,))
+
+    n_trials = 10000
+    rms = RunningMeanStd()
+    tfrms = TfRunningMeanStd()
+
+    tic1 = time.time()
+    for _ in range(n_trials):
+        rms.update(x)
+
+    tic2 = time.time()
+    for _ in range(n_trials):
+        tfrms.update(x)
+
+    tic3 = time.time()
+
+    print('rms update time ({} trials): {} s'.format(n_trials, tic2 - tic1))
+    print('tfrms update time ({} trials): {} s'.format(n_trials, tic3 - tic2))
+
+
+    tic1 = time.time()
+    for _ in range(n_trials):
+        z1 = rms.mean
+
+    tic2 = time.time()
+    for _ in range(n_trials):
+        z2 = tfrms.mean
+
+    assert z1 == z2
+
+    tic3 = time.time()
+
+    print('rms get mean time ({} trials): {} s'.format(n_trials, tic2 - tic1))
+    print('tfrms get mean time ({} trials): {} s'.format(n_trials, tic3 - tic2))
+
+
+
+    '''
+    options = tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE) #pylint: disable=E1101
+    run_metadata = tf.RunMetadata()
+    profile_opts = dict(options=options, run_metadata=run_metadata)
+
+
+
+    from tensorflow.python.client import timeline
+    fetched_timeline = timeline.Timeline(run_metadata.step_stats) #pylint: disable=E1101
+    chrome_trace = fetched_timeline.generate_chrome_trace_format()
+    outfile = '/tmp/timeline.json'
+    with open(outfile, 'wt') as f:
+        f.write(chrome_trace)
+    print('Successfully saved profile to {}. Exiting.'.format(outfile))
+    exit(0)
+    '''
+
+
+
+if __name__ == '__main__':
+   profile_tf_runningmeanstd()

baselines/common/test_mpi_util.py

@@ -0,0 +1,29 @@
+from baselines.common import mpi_util
+from baselines import logger
+from baselines.common.tests.test_with_mpi import with_mpi
+try:
+    from mpi4py import MPI
+except ImportError:
+    MPI = None
+
+@with_mpi()
+def test_mpi_weighted_mean():
+    comm = MPI.COMM_WORLD
+    with logger.scoped_configure(comm=comm):
+        if comm.rank == 0:
+            name2valcount = {'a' : (10, 2), 'b' : (20,3)}
+        elif comm.rank == 1:
+            name2valcount = {'a' : (19, 1), 'c' : (42,3)}
+        else:
+            raise NotImplementedError
+        d = mpi_util.mpi_weighted_mean(comm, name2valcount)
+        correctval = {'a' : (10 * 2 + 19) / 3.0, 'b' : 20, 'c' : 42}
+        if comm.rank == 0:
+            assert d == correctval, '{} != {}'.format(d, correctval)
+
+        for name, (val, count) in name2valcount.items():
+            for _ in range(count):
+                logger.logkv_mean(name, val)
+        d2 = logger.dumpkvs()
+        if comm.rank == 0:
+            assert d2 == correctval

baselines/common/tests/__init__.py

@@ -0,0 +1,2 @@
+import os, pytest
+mark_slow = pytest.mark.skipif(not os.getenv('RUNSLOW'), reason='slow')

baselines/common/tests/envs/fixed_sequence_env.py

@@ -0,0 +1,43 @@
+import numpy as np
+from gym import Env
+from gym.spaces import Discrete
+
+
+class FixedSequenceEnv(Env):
+    def __init__(
+            self,
+            n_actions=10,
+            episode_len=100
+    ):
+        self.action_space = Discrete(n_actions)
+        self.observation_space = Discrete(1)
+        self.np_random = np.random.RandomState(0)
+        self.episode_len = episode_len
+        self.sequence = [self.np_random.randint(0, self.action_space.n)
+            for _ in range(self.episode_len)]
+        self.time = 0
+
+
+    def reset(self):
+        self.time = 0
+        return 0
+
+    def step(self, actions):
+        rew = self._get_reward(actions)
+        self._choose_next_state()
+        done = False
+        if self.episode_len and self.time >= self.episode_len:
+            done = True
+
+        return 0, rew, done, {}
+
+    def seed(self, seed=None):
+        self.np_random.seed(seed)
+
+    def _choose_next_state(self):
+        self.time += 1
+
+    def _get_reward(self, actions):
+        return 1 if actions == self.sequence[self.time] else 0
+
+

baselines/common/tests/envs/identity_env.py

@@ -0,0 +1,90 @@
+import numpy as np
+from abc import abstractmethod
+from gym import Env
+from gym.spaces import MultiDiscrete, Discrete, Box
+from collections import deque
+
+class IdentityEnv(Env):
+    def __init__(
+            self,
+            episode_len=None,
+            delay=0,
+            zero_first_rewards=True
+    ):
+
+        self.observation_space = self.action_space
+        self.episode_len = episode_len
+        self.time = 0
+        self.delay = delay
+        self.zero_first_rewards = zero_first_rewards
+        self.q = deque(maxlen=delay+1)
+
+    def reset(self):
+        self.q.clear()
+        for _ in range(self.delay + 1):
+            self.q.append(self.action_space.sample())
+        self.time = 0
+
+        return self.q[-1]
+
+    def step(self, actions):
+        rew = self._get_reward(self.q.popleft(), actions)
+        if self.zero_first_rewards and self.time < self.delay:
+            rew = 0
+        self.q.append(self.action_space.sample())
+        self.time += 1
+        done = self.episode_len is not None and self.time >= self.episode_len
+        return self.q[-1], rew, done, {}
+
+    def seed(self, seed=None):
+        self.action_space.seed(seed)
+
+    @abstractmethod
+    def _get_reward(self, state, actions):
+        raise NotImplementedError
+
+
+class DiscreteIdentityEnv(IdentityEnv):
+    def __init__(
+            self,
+            dim,
+            episode_len=None,
+            delay=0,
+            zero_first_rewards=True
+    ):
+
+        self.action_space = Discrete(dim)
+        super().__init__(episode_len=episode_len, delay=delay, zero_first_rewards=zero_first_rewards)
+
+    def _get_reward(self, state, actions):
+        return 1 if state == actions else 0
+
+class MultiDiscreteIdentityEnv(IdentityEnv):
+    def __init__(
+            self,
+            dims,
+            episode_len=None,
+            delay=0,
+    ):
+
+        self.action_space = MultiDiscrete(dims)
+        super().__init__(episode_len=episode_len, delay=delay)
+
+    def _get_reward(self, state, actions):
+        return 1 if all(state == actions) else 0
+
+
+class BoxIdentityEnv(IdentityEnv):
+    def __init__(
+            self,
+            shape,
+            episode_len=None,
+    ):
+
+        self.action_space = Box(low=-1.0, high=1.0, shape=shape, dtype=np.float32)
+        super().__init__(episode_len=episode_len)
+
+    def _get_reward(self, state, actions):
+        diff = actions - state
+        diff = diff[:]
+        return -0.5 * np.dot(diff, diff)

baselines/common/tests/envs/identity_env_test.py

@@ -0,0 +1,36 @@
+from baselines.common.tests.envs.identity_env import DiscreteIdentityEnv
+
+
+def test_discrete_nodelay():
+    nsteps = 100
+    eplen = 50
+    env = DiscreteIdentityEnv(10, episode_len=eplen)
+    ob = env.reset()
+    for t in range(nsteps):
+        action = env.action_space.sample()
+        next_ob, rew, done, info = env.step(action)
+        assert rew == (1 if action == ob else 0)
+        if (t + 1) % eplen == 0:
+            assert done
+            next_ob = env.reset()
+        else:
+            assert not done
+        ob = next_ob
+
+def test_discrete_delay1():
+    eplen = 50
+    env = DiscreteIdentityEnv(10, episode_len=eplen, delay=1)
+    ob = env.reset()
+    prev_ob = None
+    for t in range(eplen):
+        action = env.action_space.sample()
+        next_ob, rew, done, info = env.step(action)
+        if t > 0:
+            assert rew == (1 if action == prev_ob else 0)
+        else:
+            assert rew == 0
+        prev_ob = ob
+        ob = next_ob
+        if t < eplen - 1:
+            assert not done
+    assert done

baselines/common/tests/envs/mnist_env.py

@@ -0,0 +1,71 @@
+import os.path as osp
+import numpy as np
+import tempfile
+from gym import Env
+from gym.spaces import Discrete, Box
+
+
+
+class MnistEnv(Env):
+    def __init__(
+            self,
+            episode_len=None,
+            no_images=None
+    ):
+        import filelock
+        from tensorflow.examples.tutorials.mnist import input_data
+        # we could use temporary directory for this with a context manager and
+        # TemporaryDirecotry, but then each test that uses mnist would re-download the data
+        # this way the data is not cleaned up, but we only download it once per machine
+        mnist_path = osp.join(tempfile.gettempdir(), 'MNIST_data')
+        with filelock.FileLock(mnist_path + '.lock'):
+           self.mnist = input_data.read_data_sets(mnist_path)
+
+        self.np_random = np.random.RandomState()
+
+        self.observation_space = Box(low=0.0, high=1.0, shape=(28,28,1))
+        self.action_space = Discrete(10)
+        self.episode_len = episode_len
+        self.time = 0
+        self.no_images = no_images
+
+        self.train_mode()
+        self.reset()
+
+    def reset(self):
+        self._choose_next_state()
+        self.time = 0
+
+        return self.state[0]
+
+    def step(self, actions):
+        rew = self._get_reward(actions)
+        self._choose_next_state()
+        done = False
+        if self.episode_len and self.time >= self.episode_len:
+            rew = 0
+            done = True
+
+        return self.state[0], rew, done, {}
+
+    def seed(self, seed=None):
+        self.np_random.seed(seed)
+
+    def train_mode(self):
+        self.dataset = self.mnist.train
+
+    def test_mode(self):
+        self.dataset = self.mnist.test
+
+    def _choose_next_state(self):
+        max_index = (self.no_images if self.no_images is not None else self.dataset.num_examples) - 1
+        index = self.np_random.randint(0, max_index)
+        image = self.dataset.images[index].reshape(28,28,1)*255
+        label = self.dataset.labels[index]
+        self.state = (image, label)
+        self.time += 1
+
+    def _get_reward(self, actions):
+        return 1 if self.state[1] == actions else 0
+
+

baselines/common/tests/test_cartpole.py

@@ -0,0 +1,45 @@
+import pytest
+import gym
+
+from baselines.run import get_learn_function
+from baselines.common.tests.util import reward_per_episode_test
+from baselines.common.tests import mark_slow
+
+common_kwargs = dict(
+    total_timesteps=30000,
+    network='mlp',
+    gamma=1.0,
+    seed=0,
+)
+
+learn_kwargs = {
+    'a2c' : dict(nsteps=32, value_network='copy', lr=0.05),
+    'acer': dict(value_network='copy'),
+    'acktr': dict(nsteps=32, value_network='copy', is_async=False),
+    'deepq': dict(total_timesteps=20000),
+    'ppo2': dict(value_network='copy'),
+    'trpo_mpi': {}
+}
+
+@mark_slow
+@pytest.mark.parametrize("alg", learn_kwargs.keys())
+def test_cartpole(alg):
+    '''
+    Test if the algorithm (with an mlp policy)
+    can learn to balance the cartpole
+    '''
+
+    kwargs = common_kwargs.copy()
+    kwargs.update(learn_kwargs[alg])
+
+    learn_fn = lambda e: get_learn_function(alg)(env=e, **kwargs)
+    def env_fn():
+
+        env = gym.make('CartPole-v0')
+        env.seed(0)
+        return env
+
+    reward_per_episode_test(env_fn, learn_fn, 100)
+
+if __name__ == '__main__':
+    test_cartpole('acer')

baselines/common/tests/test_doc_examples.py

@@ -0,0 +1,48 @@
+import pytest
+try:
+    import mujoco_py
+    _mujoco_present = True
+except BaseException:
+    mujoco_py = None
+    _mujoco_present = False
+
+
+@pytest.mark.skipif(
+    not _mujoco_present,
+    reason='error loading mujoco - either mujoco / mujoco key not present, or LD_LIBRARY_PATH is not pointing to mujoco library'
+)
+def test_lstm_example():
+    import tensorflow as tf
+    from baselines.common import policies, models, cmd_util
+    from baselines.common.vec_env.dummy_vec_env import DummyVecEnv
+
+    # create vectorized environment
+    venv = DummyVecEnv([lambda: cmd_util.make_mujoco_env('Reacher-v2', seed=0)])
+
+    with tf.Session() as sess:
+        # build policy based on lstm network with 128 units
+        policy = policies.build_policy(venv, models.lstm(128))(nbatch=1, nsteps=1)
+
+        # initialize tensorflow variables
+        sess.run(tf.global_variables_initializer())
+
+        # prepare environment variables
+        ob = venv.reset()
+        state = policy.initial_state
+        done = [False]
+        step_counter = 0
+
+        # run a single episode until the end (i.e. until done)
+        while True:
+            action, _, state, _ = policy.step(ob, S=state, M=done)
+            ob, reward, done, _ = venv.step(action)
+            step_counter += 1
+            if done:
+                break
+
+
+        assert step_counter > 5
+
+
+
+

baselines/common/tests/test_env_after_learn.py

@@ -0,0 +1,27 @@
+import pytest
+import gym
+import tensorflow as tf
+
+from baselines.common.vec_env.subproc_vec_env import SubprocVecEnv
+from baselines.run import get_learn_function
+from baselines.common.tf_util import make_session
+
+algos = ['a2c', 'acer', 'acktr', 'deepq', 'ppo2', 'trpo_mpi']
+
+@pytest.mark.parametrize('algo', algos)
+def test_env_after_learn(algo):
+    def make_env():
+        # acktr requires too much RAM, fails on travis
+        env = gym.make('CartPole-v1' if algo == 'acktr' else 'PongNoFrameskip-v4')
+        return env
+
+    make_session(make_default=True, graph=tf.Graph())
+    env = SubprocVecEnv([make_env])
+
+    learn = get_learn_function(algo)
+
+    # Commenting out the following line resolves the issue, though crash happens at env.reset().
+    learn(network='mlp', env=env, total_timesteps=0, load_path=None, seed=None)
+
+    env.reset()
+    env.close()

baselines/common/tests/test_fetchreach.py

@@ -0,0 +1,40 @@
+import pytest
+import gym
+
+from baselines.run import get_learn_function
+from baselines.common.tests.util import reward_per_episode_test
+from baselines.common.tests import mark_slow
+
+pytest.importorskip('mujoco_py')
+
+common_kwargs = dict(
+    network='mlp',
+    seed=0,
+)
+
+learn_kwargs = {
+    'her': dict(total_timesteps=2000)
+}
+
+@mark_slow
+@pytest.mark.parametrize("alg", learn_kwargs.keys())
+def test_fetchreach(alg):
+    '''
+    Test if the algorithm (with an mlp policy)
+    can learn the FetchReach task
+    '''
+
+    kwargs = common_kwargs.copy()
+    kwargs.update(learn_kwargs[alg])
+
+    learn_fn = lambda e: get_learn_function(alg)(env=e, **kwargs)
+    def env_fn():
+
+        env = gym.make('FetchReach-v1')
+        env.seed(0)
+        return env
+
+    reward_per_episode_test(env_fn, learn_fn, -15)
+
+if __name__ == '__main__':
+    test_fetchreach('her')

baselines/common/tests/test_fixed_sequence.py

@@ -0,0 +1,52 @@
+import pytest
+from baselines.common.tests.envs.fixed_sequence_env import FixedSequenceEnv
+
+from baselines.common.tests.util import simple_test
+from baselines.run import get_learn_function
+from baselines.common.tests import mark_slow
+
+
+common_kwargs = dict(
+    seed=0,
+    total_timesteps=50000,
+)
+
+learn_kwargs = {
+    'a2c': {},
+    'ppo2': dict(nsteps=10, ent_coef=0.0, nminibatches=1),
+    # TODO enable sequential models for trpo_mpi (proper handling of nbatch and nsteps)
+    # github issue: https://github.com/openai/baselines/issues/188
+    # 'trpo_mpi': lambda e, p: trpo_mpi.learn(policy_fn=p(env=e), env=e, max_timesteps=30000, timesteps_per_batch=100, cg_iters=10, gamma=0.9, lam=1.0, max_kl=0.001)
+}
+
+
+alg_list = learn_kwargs.keys()
+rnn_list = ['lstm']
+
+@mark_slow
+@pytest.mark.parametrize("alg", alg_list)
+@pytest.mark.parametrize("rnn", rnn_list)
+def test_fixed_sequence(alg, rnn):
+    '''
+    Test if the algorithm (with a given policy)
+    can learn an identity transformation (i.e. return observation as an action)
+    '''
+
+    kwargs = learn_kwargs[alg]
+    kwargs.update(common_kwargs)
+
+    env_fn = lambda: FixedSequenceEnv(n_actions=10, episode_len=5)
+    learn = lambda e: get_learn_function(alg)(
+        env=e,
+        network=rnn,
+        **kwargs
+    )
+
+    simple_test(env_fn, learn, 0.7)
+
+
+if __name__ == '__main__':
+    test_fixed_sequence('ppo2', 'lstm')
+
+
+

baselines/common/tests/test_identity.py

@@ -0,0 +1,76 @@
+import pytest
+from baselines.common.tests.envs.identity_env import DiscreteIdentityEnv, BoxIdentityEnv, MultiDiscreteIdentityEnv
+from baselines.run import get_learn_function
+from baselines.common.tests.util import simple_test
+from baselines.common.tests import mark_slow
+
+common_kwargs = dict(
+    total_timesteps=30000,
+    network='mlp',
+    gamma=0.9,
+    seed=0,
+)
+
+learn_kwargs = {
+    'a2c' : {},
+    'acktr': {},
+    'deepq': {},
+    'ddpg': dict(layer_norm=True),
+    'ppo2': dict(lr=1e-3, nsteps=64, ent_coef=0.0),
+    'trpo_mpi': dict(timesteps_per_batch=100, cg_iters=10, gamma=0.9, lam=1.0, max_kl=0.01)
+}
+
+
+algos_disc = ['a2c', 'acktr', 'deepq', 'ppo2', 'trpo_mpi']
+algos_multidisc = ['a2c', 'acktr', 'ppo2', 'trpo_mpi']
+algos_cont = ['a2c', 'acktr', 'ddpg',  'ppo2', 'trpo_mpi']
+
+@mark_slow
+@pytest.mark.parametrize("alg", algos_disc)
+def test_discrete_identity(alg):
+    '''
+    Test if the algorithm (with an mlp policy)
+    can learn an identity transformation (i.e. return observation as an action)
+    '''
+
+    kwargs = learn_kwargs[alg]
+    kwargs.update(common_kwargs)
+
+    learn_fn = lambda e: get_learn_function(alg)(env=e, **kwargs)
+    env_fn = lambda: DiscreteIdentityEnv(10, episode_len=100)
+    simple_test(env_fn, learn_fn, 0.9)
+
+@mark_slow
+@pytest.mark.parametrize("alg", algos_multidisc)
+def test_multidiscrete_identity(alg):
+    '''
+    Test if the algorithm (with an mlp policy)
+    can learn an identity transformation (i.e. return observation as an action)
+    '''
+
+    kwargs = learn_kwargs[alg]
+    kwargs.update(common_kwargs)
+
+    learn_fn = lambda e: get_learn_function(alg)(env=e, **kwargs)
+    env_fn = lambda: MultiDiscreteIdentityEnv((3,3), episode_len=100)
+    simple_test(env_fn, learn_fn, 0.9)
+
+@mark_slow
+@pytest.mark.parametrize("alg", algos_cont)
+def test_continuous_identity(alg):
+    '''
+    Test if the algorithm (with an mlp policy)
+    can learn an identity transformation (i.e. return observation as an action)
+    to a required precision
+    '''
+
+    kwargs = learn_kwargs[alg]
+    kwargs.update(common_kwargs)
+    learn_fn = lambda e: get_learn_function(alg)(env=e, **kwargs)
+
+    env_fn = lambda: BoxIdentityEnv((1,), episode_len=100)
+    simple_test(env_fn, learn_fn, -0.1)
+
+if __name__ == '__main__':
+    test_multidiscrete_identity('acktr')
+

baselines/common/tests/test_mnist.py

@@ -0,0 +1,49 @@
+import pytest
+
+# from baselines.acer import acer_simple as acer
+from baselines.common.tests.envs.mnist_env import MnistEnv
+from baselines.common.tests.util import simple_test
+from baselines.run import get_learn_function
+from baselines.common.tests import mark_slow
+
+# TODO investigate a2c and ppo2 failures - is it due to bad hyperparameters for this problem?
+# GitHub issue https://github.com/openai/baselines/issues/189
+common_kwargs = {
+    'seed': 0,
+    'network':'cnn',
+    'gamma':0.9,
+    'pad':'SAME'
+}
+
+learn_args = {
+    'a2c': dict(total_timesteps=50000),
+    'acer': dict(total_timesteps=20000),
+    'deepq': dict(total_timesteps=5000),
+    'acktr': dict(total_timesteps=30000),
+    'ppo2': dict(total_timesteps=50000, lr=1e-3, nsteps=128, ent_coef=0.0),
+    'trpo_mpi': dict(total_timesteps=80000, timesteps_per_batch=100, cg_iters=10, lam=1.0, max_kl=0.001)
+}
+
+
+#tests pass, but are too slow on travis. Same algorithms are covered
+# by other tests with less compute-hungry nn's and by benchmarks
+@pytest.mark.skip
+@mark_slow
+@pytest.mark.parametrize("alg", learn_args.keys())
+def test_mnist(alg):
+    '''
+    Test if the algorithm can learn to classify MNIST digits.
+    Uses CNN policy.
+    '''
+
+    learn_kwargs = learn_args[alg]
+    learn_kwargs.update(common_kwargs)
+
+    learn = get_learn_function(alg)
+    learn_fn = lambda e: learn(env=e, **learn_kwargs)
+    env_fn = lambda: MnistEnv(episode_len=100)
+
+    simple_test(env_fn, learn_fn, 0.6)
+
+if __name__ == '__main__':
+    test_mnist('acer')

baselines/common/tests/test_plot_util.py

@@ -0,0 +1,17 @@
+# smoke tests of plot_util
+from baselines.common import plot_util as pu
+from baselines.common.tests.util import smoketest
+
+
+def test_plot_util():
+    nruns = 4
+    logdirs = [smoketest('--alg=ppo2 --env=CartPole-v0 --num_timesteps=10000') for _ in range(nruns)]
+    data = pu.load_results(logdirs)
+    assert len(data) == 4
+
+    _, axes = pu.plot_results(data[:1]); assert len(axes) == 1
+    _, axes = pu.plot_results(data, tiling='vertical'); assert axes.shape==(4,1)
+    _, axes = pu.plot_results(data, tiling='horizontal'); assert axes.shape==(1,4)
+    _, axes = pu.plot_results(data, tiling='symmetric'); assert axes.shape==(2,2)
+    _, axes = pu.plot_results(data, split_fn=lambda _: ''); assert len(axes) == 1
+

baselines/common/tests/test_schedules.py

@@ -0,0 +1,26 @@
+import numpy as np
+
+from baselines.common.schedules import ConstantSchedule, PiecewiseSchedule
+
+
+def test_piecewise_schedule():
+    ps = PiecewiseSchedule([(-5, 100), (5, 200), (10, 50), (100, 50), (200, -50)], outside_value=500)
+
+    assert np.isclose(ps.value(-10), 500)
+    assert np.isclose(ps.value(0), 150)
+    assert np.isclose(ps.value(5), 200)
+    assert np.isclose(ps.value(9), 80)
+    assert np.isclose(ps.value(50), 50)
+    assert np.isclose(ps.value(80), 50)
+    assert np.isclose(ps.value(150), 0)
+    assert np.isclose(ps.value(175), -25)
+    assert np.isclose(ps.value(201), 500)
+    assert np.isclose(ps.value(500), 500)
+
+    assert np.isclose(ps.value(200 - 1e-10), -50)
+
+
+def test_constant_schedule():
+    cs = ConstantSchedule(5)
+    for i in range(-100, 100):
+        assert np.isclose(cs.value(i), 5)

baselines/common/tests/test_segment_tree.py

@@ -0,0 +1,103 @@
+import numpy as np
+
+from baselines.common.segment_tree import SumSegmentTree, MinSegmentTree
+
+
+def test_tree_set():
+    tree = SumSegmentTree(4)
+
+    tree[2] = 1.0
+    tree[3] = 3.0
+
+    assert np.isclose(tree.sum(), 4.0)
+    assert np.isclose(tree.sum(0, 2), 0.0)
+    assert np.isclose(tree.sum(0, 3), 1.0)
+    assert np.isclose(tree.sum(2, 3), 1.0)
+    assert np.isclose(tree.sum(2, -1), 1.0)
+    assert np.isclose(tree.sum(2, 4), 4.0)
+
+
+def test_tree_set_overlap():
+    tree = SumSegmentTree(4)
+
+    tree[2] = 1.0
+    tree[2] = 3.0
+
+    assert np.isclose(tree.sum(), 3.0)
+    assert np.isclose(tree.sum(2, 3), 3.0)
+    assert np.isclose(tree.sum(2, -1), 3.0)
+    assert np.isclose(tree.sum(2, 4), 3.0)
+    assert np.isclose(tree.sum(1, 2), 0.0)
+
+
+def test_prefixsum_idx():
+    tree = SumSegmentTree(4)
+
+    tree[2] = 1.0
+    tree[3] = 3.0
+
+    assert tree.find_prefixsum_idx(0.0) == 2
+    assert tree.find_prefixsum_idx(0.5) == 2
+    assert tree.find_prefixsum_idx(0.99) == 2
+    assert tree.find_prefixsum_idx(1.01) == 3
+    assert tree.find_prefixsum_idx(3.00) == 3
+    assert tree.find_prefixsum_idx(4.00) == 3
+
+
+def test_prefixsum_idx2():
+    tree = SumSegmentTree(4)
+
+    tree[0] = 0.5
+    tree[1] = 1.0
+    tree[2] = 1.0
+    tree[3] = 3.0
+
+    assert tree.find_prefixsum_idx(0.00) == 0
+    assert tree.find_prefixsum_idx(0.55) == 1
+    assert tree.find_prefixsum_idx(0.99) == 1
+    assert tree.find_prefixsum_idx(1.51) == 2
+    assert tree.find_prefixsum_idx(3.00) == 3
+    assert tree.find_prefixsum_idx(5.50) == 3
+
+
+def test_max_interval_tree():
+    tree = MinSegmentTree(4)
+
+    tree[0] = 1.0
+    tree[2] = 0.5
+    tree[3] = 3.0
+
+    assert np.isclose(tree.min(), 0.5)
+    assert np.isclose(tree.min(0, 2), 1.0)
+    assert np.isclose(tree.min(0, 3), 0.5)
+    assert np.isclose(tree.min(0, -1), 0.5)
+    assert np.isclose(tree.min(2, 4), 0.5)
+    assert np.isclose(tree.min(3, 4), 3.0)
+
+    tree[2] = 0.7
+
+    assert np.isclose(tree.min(), 0.7)
+    assert np.isclose(tree.min(0, 2), 1.0)
+    assert np.isclose(tree.min(0, 3), 0.7)
+    assert np.isclose(tree.min(0, -1), 0.7)
+    assert np.isclose(tree.min(2, 4), 0.7)
+    assert np.isclose(tree.min(3, 4), 3.0)
+
+    tree[2] = 4.0
+
+    assert np.isclose(tree.min(), 1.0)
+    assert np.isclose(tree.min(0, 2), 1.0)
+    assert np.isclose(tree.min(0, 3), 1.0)
+    assert np.isclose(tree.min(0, -1), 1.0)
+    assert np.isclose(tree.min(2, 4), 3.0)
+    assert np.isclose(tree.min(2, 3), 4.0)
+    assert np.isclose(tree.min(2, -1), 4.0)
+    assert np.isclose(tree.min(3, 4), 3.0)
+
+
+if __name__ == '__main__':
+    test_tree_set()
+    test_tree_set_overlap()
+    test_prefixsum_idx()
+    test_prefixsum_idx2()
+    test_max_interval_tree()

baselines/common/tests/test_serialization.py

@@ -0,0 +1,139 @@
+import os
+import gym
+import tempfile
+import pytest
+import tensorflow as tf
+import numpy as np
+
+from baselines.common.tests.envs.mnist_env import MnistEnv
+from baselines.common.vec_env.dummy_vec_env import DummyVecEnv
+from baselines.run import get_learn_function
+from baselines.common.tf_util import make_session, get_session
+
+from functools import partial
+
+
+learn_kwargs = {
+    'deepq': {},
+    'a2c': {},
+    'acktr': {},
+    'acer': {},
+    'ppo2': {'nminibatches': 1, 'nsteps': 10},
+    'trpo_mpi': {},
+}
+
+network_kwargs = {
+    'mlp': {},
+    'cnn': {'pad': 'SAME'},
+    'lstm': {},
+    'cnn_lnlstm': {'pad': 'SAME'}
+}
+
+
+@pytest.mark.parametrize("learn_fn", learn_kwargs.keys())
+@pytest.mark.parametrize("network_fn", network_kwargs.keys())
+def test_serialization(learn_fn, network_fn):
+    '''
+    Test if the trained model can be serialized
+    '''
+
+
+    if network_fn.endswith('lstm') and learn_fn in ['acer', 'acktr', 'trpo_mpi', 'deepq']:
+            # TODO make acktr work with recurrent policies
+            # and test
+            # github issue: https://github.com/openai/baselines/issues/660
+            return
+
+    def make_env():
+        env = MnistEnv(episode_len=100)
+        env.seed(10)
+        return env
+
+    env = DummyVecEnv([make_env])
+    ob = env.reset().copy()
+    learn = get_learn_function(learn_fn)
+
+    kwargs = {}
+    kwargs.update(network_kwargs[network_fn])
+    kwargs.update(learn_kwargs[learn_fn])
+
+
+    learn = partial(learn, env=env, network=network_fn, seed=0, **kwargs)
+
+    with tempfile.TemporaryDirectory() as td:
+        model_path = os.path.join(td, 'serialization_test_model')
+
+        with tf.Graph().as_default(), make_session().as_default():
+            model = learn(total_timesteps=100)
+            model.save(model_path)
+            mean1, std1 = _get_action_stats(model, ob)
+            variables_dict1 = _serialize_variables()
+
+        with tf.Graph().as_default(), make_session().as_default():
+            model = learn(total_timesteps=0, load_path=model_path)
+            mean2, std2 = _get_action_stats(model, ob)
+            variables_dict2 = _serialize_variables()
+
+        for k, v in variables_dict1.items():
+            np.testing.assert_allclose(v, variables_dict2[k], atol=0.01,
+                err_msg='saved and loaded variable {} value mismatch'.format(k))
+
+        np.testing.assert_allclose(mean1, mean2, atol=0.5)
+        np.testing.assert_allclose(std1, std2, atol=0.5)
+
+
+@pytest.mark.parametrize("learn_fn", learn_kwargs.keys())
+@pytest.mark.parametrize("network_fn", ['mlp'])
+def test_coexistence(learn_fn, network_fn):
+    '''
+    Test if more than one model can exist at a time
+    '''
+
+    if learn_fn == 'deepq':
+            # TODO enable multiple DQN models to be useable at the same time
+            # github issue https://github.com/openai/baselines/issues/656
+            return
+
+    if network_fn.endswith('lstm') and learn_fn in ['acktr', 'trpo_mpi', 'deepq']:
+            # TODO make acktr work with recurrent policies
+            # and test
+            # github issue: https://github.com/openai/baselines/issues/660
+            return
+
+    env = DummyVecEnv([lambda: gym.make('CartPole-v0')])
+    learn = get_learn_function(learn_fn)
+
+    kwargs = {}
+    kwargs.update(network_kwargs[network_fn])
+    kwargs.update(learn_kwargs[learn_fn])
+
+    learn =  partial(learn, env=env, network=network_fn, total_timesteps=0, **kwargs)
+    make_session(make_default=True, graph=tf.Graph())
+    model1 = learn(seed=1)
+    make_session(make_default=True, graph=tf.Graph())
+    model2 = learn(seed=2)
+
+    model1.step(env.observation_space.sample())
+    model2.step(env.observation_space.sample())
+
+
+
+def _serialize_variables():
+    sess = get_session()
+    variables = tf.trainable_variables()
+    values = sess.run(variables)
+    return {var.name: value for var, value in zip(variables, values)}
+
+
+def _get_action_stats(model, ob):
+    ntrials = 1000
+    if model.initial_state is None or model.initial_state == []:
+        actions = np.array([model.step(ob)[0] for _ in range(ntrials)])
+    else:
+        actions = np.array([model.step(ob, S=model.initial_state, M=[False])[0] for _ in range(ntrials)])
+
+    mean = np.mean(actions, axis=0)
+    std = np.std(actions, axis=0)
+
+    return mean, std
+

baselines/common/tests/test_tf_util.py

@@ -0,0 +1,42 @@
+# tests for tf_util
+import tensorflow as tf
+from baselines.common.tf_util import (
+    function,
+    initialize,
+    single_threaded_session
+)
+
+
+def test_function():
+    with tf.Graph().as_default():
+        x = tf.placeholder(tf.int32, (), name="x")
+        y = tf.placeholder(tf.int32, (), name="y")
+        z = 3 * x + 2 * y
+        lin = function([x, y], z, givens={y: 0})
+
+        with single_threaded_session():
+            initialize()
+
+            assert lin(2) == 6
+            assert lin(x=3) == 9
+            assert lin(2, 2) == 10
+            assert lin(x=2, y=3) == 12
+
+
+def test_multikwargs():
+    with tf.Graph().as_default():
+        x = tf.placeholder(tf.int32, (), name="x")
+        with tf.variable_scope("other"):
+            x2 = tf.placeholder(tf.int32, (), name="x")
+        z = 3 * x + 2 * x2
+
+        lin = function([x, x2], z, givens={x2: 0})
+        with single_threaded_session():
+            initialize()
+            assert lin(2) == 6
+            assert lin(2, 2) == 10
+
+
+if __name__ == '__main__':
+    test_function()
+    test_multikwargs()

baselines/common/tests/util.py

@@ -0,0 +1,92 @@
+import tensorflow as tf
+import numpy as np
+from baselines.common.vec_env.dummy_vec_env import DummyVecEnv
+
+N_TRIALS = 10000
+N_EPISODES = 100
+
+_sess_config = tf.ConfigProto(
+    allow_soft_placement=True,
+    intra_op_parallelism_threads=1,
+    inter_op_parallelism_threads=1
+)
+
+def simple_test(env_fn, learn_fn, min_reward_fraction, n_trials=N_TRIALS):
+    def seeded_env_fn():
+        env = env_fn()
+        env.seed(0)
+        return env
+
+    np.random.seed(0)
+    env = DummyVecEnv([seeded_env_fn])
+    with tf.Graph().as_default(), tf.Session(config=_sess_config).as_default():
+        tf.set_random_seed(0)
+        model = learn_fn(env)
+        sum_rew = 0
+        done = True
+        for i in range(n_trials):
+            if done:
+                obs = env.reset()
+                state = model.initial_state
+            if state is not None:
+                a, v, state, _ = model.step(obs, S=state, M=[False])
+            else:
+                a, v, _, _ = model.step(obs)
+            obs, rew, done, _ = env.step(a)
+            sum_rew += float(rew)
+        print("Reward in {} trials is {}".format(n_trials, sum_rew))
+        assert sum_rew > min_reward_fraction * n_trials, \
+            'sum of rewards {} is less than {} of the total number of trials {}'.format(sum_rew, min_reward_fraction, n_trials)
+
+def reward_per_episode_test(env_fn, learn_fn, min_avg_reward, n_trials=N_EPISODES):
+    env = DummyVecEnv([env_fn])
+    with tf.Graph().as_default(), tf.Session(config=_sess_config).as_default():
+        model = learn_fn(env)
+        N_TRIALS = 100
+        observations, actions, rewards = rollout(env, model, N_TRIALS)
+        rewards = [sum(r) for r in rewards]
+        avg_rew = sum(rewards) / N_TRIALS
+        print("Average reward in {} episodes is {}".format(n_trials, avg_rew))
+        assert avg_rew > min_avg_reward, \
+            'average reward in {} episodes ({}) is less than {}'.format(n_trials, avg_rew, min_avg_reward)
+
+def rollout(env, model, n_trials):
+    rewards = []
+    actions = []
+    observations = []
+    for i in range(n_trials):
+        obs = env.reset()
+        state = model.initial_state if hasattr(model, 'initial_state') else None
+        episode_rew = []
+        episode_actions = []
+        episode_obs = []
+        while True:
+            if state is not None:
+                a, v, state, _ = model.step(obs, S=state, M=[False])
+            else:
+                a,v, _, _ = model.step(obs)
+
+            obs, rew, done, _ = env.step(a)
+            episode_rew.append(rew)
+            episode_actions.append(a)
+            episode_obs.append(obs)
+            if done:
+                break
+        rewards.append(episode_rew)
+        actions.append(episode_actions)
+        observations.append(episode_obs)
+    return observations, actions, rewards
+
+
+def smoketest(argstr, **kwargs):
+    import tempfile
+    import subprocess
+    import os
+    argstr = 'python -m baselines.run ' + argstr
+    for key, value in kwargs:
+        argstr += ' --{}={}'.format(key, value)
+    tempdir = tempfile.mkdtemp()
+    env = os.environ.copy()
+    env['OPENAI_LOGDIR'] = tempdir
+    subprocess.run(argstr.split(' '), env=env)
+    return tempdir

baselines/common/tf_util.py

@@ -1,6 +1,10 @@
 import numpy as np
 import tensorflow as tf  # pylint: ignore-module
 import copy
+import os
+import functools
+import collections
+import multiprocessing
 
 def switch(condition, then_expression, else_expression):
     """Switches between two operations depending on a scalar value (int or bool).
@@ -40,6 +44,52 @@ def huber_loss(x, delta=1.0):
         delta * (tf.abs(x) - 0.5 * delta)
     )
 
+# ================================================================
+# Global session
+# ================================================================
+
+def get_session(config=None):
+    """Get default session or create one with a given config"""
+    sess = tf.get_default_session()
+    if sess is None:
+        sess = make_session(config=config, make_default=True)
+    return sess
+
+def make_session(config=None, num_cpu=None, make_default=False, graph=None):
+    """Returns a session that will use <num_cpu> CPU's only"""
+    if num_cpu is None:
+        num_cpu = int(os.getenv('RCALL_NUM_CPU', multiprocessing.cpu_count()))
+    if config is None:
+        config = tf.ConfigProto(
+            allow_soft_placement=True,
+            inter_op_parallelism_threads=num_cpu,
+            intra_op_parallelism_threads=num_cpu)
+        config.gpu_options.allow_growth = True
+
+    if make_default:
+        return tf.InteractiveSession(config=config, graph=graph)
+    else:
+        return tf.Session(config=config, graph=graph)
+
+def single_threaded_session():
+    """Returns a session which will only use a single CPU"""
+    return make_session(num_cpu=1)
+
+def in_session(f):
+    @functools.wraps(f)
+    def newfunc(*args, **kwargs):
+        with tf.Session():
+            f(*args, **kwargs)
+    return newfunc
+
+ALREADY_INITIALIZED = set()
+
+def initialize():
+    """Initialize all the uninitialized variables in the global scope."""
+    new_variables = set(tf.global_variables()) - ALREADY_INITIALIZED
+    get_session().run(tf.variables_initializer(new_variables))
+    ALREADY_INITIALIZED.update(new_variables)
+
 # ================================================================
 # Model components
 # ================================================================
@@ -80,6 +130,87 @@ def conv2d(x, num_filters, name, filter_size=(3, 3), stride=(1, 1), pad="SAME",
 
         return tf.nn.conv2d(x, w, stride_shape, pad) + b
 
+# ================================================================
+# Theano-like Function
+# ================================================================
+
+def function(inputs, outputs, updates=None, givens=None):
+    """Just like Theano function. Take a bunch of tensorflow placeholders and expressions
+    computed based on those placeholders and produces f(inputs) -> outputs. Function f takes
+    values to be fed to the input's placeholders and produces the values of the expressions
+    in outputs.
+
+    Input values can be passed in the same order as inputs or can be provided as kwargs based
+    on placeholder name (passed to constructor or accessible via placeholder.op.name).
+
+    Example:
+        x = tf.placeholder(tf.int32, (), name="x")
+        y = tf.placeholder(tf.int32, (), name="y")
+        z = 3 * x + 2 * y
+        lin = function([x, y], z, givens={y: 0})
+
+        with single_threaded_session():
+            initialize()
+
+            assert lin(2) == 6
+            assert lin(x=3) == 9
+            assert lin(2, 2) == 10
+            assert lin(x=2, y=3) == 12
+
+    Parameters
+    ----------
+    inputs: [tf.placeholder, tf.constant, or object with make_feed_dict method]
+        list of input arguments
+    outputs: [tf.Variable] or tf.Variable
+        list of outputs or a single output to be returned from function. Returned
+        value will also have the same shape.
+    updates: [tf.Operation] or tf.Operation
+        list of update functions or single update function that will be run whenever
+        the function is called. The return is ignored.
+
+    """
+    if isinstance(outputs, list):
+        return _Function(inputs, outputs, updates, givens=givens)
+    elif isinstance(outputs, (dict, collections.OrderedDict)):
+        f = _Function(inputs, outputs.values(), updates, givens=givens)
+        return lambda *args, **kwargs: type(outputs)(zip(outputs.keys(), f(*args, **kwargs)))
+    else:
+        f = _Function(inputs, [outputs], updates, givens=givens)
+        return lambda *args, **kwargs: f(*args, **kwargs)[0]
+
+
+class _Function(object):
+    def __init__(self, inputs, outputs, updates, givens):
+        for inpt in inputs:
+            if not hasattr(inpt, 'make_feed_dict') and not (type(inpt) is tf.Tensor and len(inpt.op.inputs) == 0):
+                assert False, "inputs should all be placeholders, constants, or have a make_feed_dict method"
+        self.inputs = inputs
+        self.input_names = {inp.name.split("/")[-1].split(":")[0]: inp for inp in inputs}
+        updates = updates or []
+        self.update_group = tf.group(*updates)
+        self.outputs_update = list(outputs) + [self.update_group]
+        self.givens = {} if givens is None else givens
+
+    def _feed_input(self, feed_dict, inpt, value):
+        if hasattr(inpt, 'make_feed_dict'):
+            feed_dict.update(inpt.make_feed_dict(value))
+        else:
+            feed_dict[inpt] = adjust_shape(inpt, value)
+
+    def __call__(self, *args, **kwargs):
+        assert len(args) + len(kwargs) <= len(self.inputs), "Too many arguments provided"
+        feed_dict = {}
+        # Update feed dict with givens.
+        for inpt in self.givens:
+            feed_dict[inpt] = adjust_shape(inpt, feed_dict.get(inpt, self.givens[inpt]))
+        # Update the args
+        for inpt, value in zip(self.inputs, args):
+            self._feed_input(feed_dict, inpt, value)
+        for inpt_name, value in kwargs.items():
+            self._feed_input(feed_dict, self.input_names[inpt_name], value)
+        results = get_session().run(self.outputs_update, feed_dict=feed_dict)[:-1]
+        return results
+
 # ================================================================
 # Flat vectors
 # ================================================================
@@ -96,7 +227,8 @@ def numel(x):
 def intprod(x):
     return int(np.prod(x))
 
-def flatgrad(grads, var_list, clip_norm=None):
+def flatgrad(loss, var_list, clip_norm=None):
+    grads = tf.gradients(loss, var_list)
     if clip_norm is not None:
         grads = [tf.clip_by_norm(grad, clip_norm=clip_norm) for grad in grads]
     return tf.concat(axis=0, values=[
@@ -106,40 +238,151 @@ def flatgrad(grads, var_list, clip_norm=None):
 
 class SetFromFlat(object):
     def __init__(self, var_list, dtype=tf.float32):
-        self.shapes = list(map(var_shape, var_list))
-        self.total_size = np.sum([intprod(shape) for shape in self.shapes])
-        self.var_list = var_list
+        assigns = []
+        shapes = list(map(var_shape, var_list))
+        total_size = np.sum([intprod(shape) for shape in shapes])
+
+        self.theta = theta = tf.placeholder(dtype, [total_size])
+        start = 0
+        assigns = []
+        for (shape, v) in zip(shapes, var_list):
+            size = intprod(shape)
+            assigns.append(tf.assign(v, tf.reshape(theta[start:start + size], shape)))
+            start += size
+        self.op = tf.group(*assigns)
 
     def __call__(self, theta):
-        start = 0
-        for (shape, v) in zip(self.shapes, self.var_list):
-            size = intprod(shape)
-            v.assign(tf.reshape(theta[start:start + size], shape))
-            start += size
+        tf.get_default_session().run(self.op, feed_dict={self.theta: theta})
 
 class GetFlat(object):
     def __init__(self, var_list):
-        self.var_list = var_list
+        self.op = tf.concat(axis=0, values=[tf.reshape(v, [numel(v)]) for v in var_list])
 
     def __call__(self):
-        return tf.concat(axis=0, values=[tf.reshape(v, [numel(v)]) for v in self.var_list]).numpy()
+        return tf.get_default_session().run(self.op)
 
 def flattenallbut0(x):
     return tf.reshape(x, [-1, intprod(x.get_shape().as_list()[1:])])
 
+# =============================================================
+# TF placeholders management
+# ============================================================
+
+_PLACEHOLDER_CACHE = {}  # name -> (placeholder, dtype, shape)
+
+def get_placeholder(name, dtype, shape):
+    if name in _PLACEHOLDER_CACHE:
+        out, dtype1, shape1 = _PLACEHOLDER_CACHE[name]
+        if out.graph == tf.get_default_graph():
+            assert dtype1 == dtype and shape1 == shape, \
+                'Placeholder with name {} has already been registered and has shape {}, different from requested {}'.format(name, shape1, shape)
+            return out
+
+    out = tf.placeholder(dtype=dtype, shape=shape, name=name)
+    _PLACEHOLDER_CACHE[name] = (out, dtype, shape)
+    return out
+
+def get_placeholder_cached(name):
+    return _PLACEHOLDER_CACHE[name][0]
+
+
 
 # ================================================================
-# Shape adjustment for feeding into tf tensors
+# Diagnostics
 # ================================================================
-def adjust_shape(input_tensor, data):
+
+def display_var_info(vars):
+    from baselines import logger
+    count_params = 0
+    for v in vars:
+        name = v.name
+        if "/Adam" in name or "beta1_power" in name or "beta2_power" in name: continue
+        v_params = np.prod(v.shape.as_list())
+        count_params += v_params
+        if "/b:" in name or "/bias" in name: continue    # Wx+b, bias is not interesting to look at => count params, but not print
+        logger.info("   %s%s %i params %s" % (name, " "*(55-len(name)), v_params, str(v.shape)))
+
+    logger.info("Total model parameters: %0.2f million" % (count_params*1e-6))
+
+
+def get_available_gpus(session_config=None):
+    # based on recipe from https://stackoverflow.com/a/38580201
+
+    # Unless we allocate a session here, subsequent attempts to create one
+    # will ignore our custom config (in particular, allow_growth=True will have
+    # no effect).
+    if session_config is None:
+        session_config = get_session()._config
+
+    from tensorflow.python.client import device_lib
+    local_device_protos = device_lib.list_local_devices(session_config)
+    return [x.name for x in local_device_protos if x.device_type == 'GPU']
+
+# ================================================================
+# Saving variables
+# ================================================================
+
+def load_state(fname, sess=None):
+    from baselines import logger
+    logger.warn('load_state method is deprecated, please use load_variables instead')
+    sess = sess or get_session()
+    saver = tf.train.Saver()
+    saver.restore(tf.get_default_session(), fname)
+
+def save_state(fname, sess=None):
+    from baselines import logger
+    logger.warn('save_state method is deprecated, please use save_variables instead')
+    sess = sess or get_session()
+    dirname = os.path.dirname(fname)
+    if any(dirname):
+        os.makedirs(dirname, exist_ok=True)
+    saver = tf.train.Saver()
+    saver.save(tf.get_default_session(), fname)
+
+# The methods above and below are clearly doing the same thing, and in a rather similar way
+# TODO: ensure there is no subtle differences and remove one
+
+def save_variables(save_path, variables=None, sess=None):
+    import joblib
+    sess = sess or get_session()
+    variables = variables or tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES)
+
+    ps = sess.run(variables)
+    save_dict = {v.name: value for v, value in zip(variables, ps)}
+    dirname = os.path.dirname(save_path)
+    if any(dirname):
+        os.makedirs(dirname, exist_ok=True)
+    joblib.dump(save_dict, save_path)
+
+def load_variables(load_path, variables=None, sess=None):
+    import joblib
+    sess = sess or get_session()
+    variables = variables or tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES)
+
+    loaded_params = joblib.load(os.path.expanduser(load_path))
+    restores = []
+    if isinstance(loaded_params, list):
+        assert len(loaded_params) == len(variables), 'number of variables loaded mismatches len(variables)'
+        for d, v in zip(loaded_params, variables):
+            restores.append(v.assign(d))
+    else:
+        for v in variables:
+            restores.append(v.assign(loaded_params[v.name]))
+
+    sess.run(restores)
+
+# ================================================================
+# Shape adjustment for feeding into tf placeholders
+# ================================================================
+def adjust_shape(placeholder, data):
     '''
-    adjust shape of the data to the shape of the tensor if possible.
+    adjust shape of the data to the shape of the placeholder if possible.
     If shape is incompatible, AssertionError is thrown
 
     Parameters:
-        input_tensor    tensorflow input tensor
+        placeholder     tensorflow input placeholder
 
-        data            input data to be (potentially) reshaped to be fed into input
+        data            input data to be (potentially) reshaped to be fed into placeholder
 
     Returns:
         reshaped data
@@ -150,23 +393,24 @@ def adjust_shape(input_tensor, data):
     if isinstance(data, list):
         data = np.array(data)
 
-    input_shape = [x or -1 for x in input_tensor.shape.as_list()]
+    placeholder_shape = [x or -1 for x in placeholder.shape.as_list()]
 
-    assert _check_shape(input_shape, data.shape), \
-        'Shape of data {} is not compatible with shape of the input {}'.format(data.shape, input_shape)
+    assert _check_shape(placeholder_shape, data.shape), \
+        'Shape of data {} is not compatible with shape of the placeholder {}'.format(data.shape, placeholder_shape)
 
-    return np.reshape(data, input_shape)
+    return np.reshape(data, placeholder_shape)
 
 
-def _check_shape(input_shape, data_shape):
+def _check_shape(placeholder_shape, data_shape):
     ''' check if two shapes are compatible (i.e. differ only by dimensions of size 1, or by the batch dimension)'''
 
-    squeezed_input_shape = _squeeze_shape(input_shape)
+    return True
+    squeezed_placeholder_shape = _squeeze_shape(placeholder_shape)
     squeezed_data_shape = _squeeze_shape(data_shape)
 
     for i, s_data in enumerate(squeezed_data_shape):
-        s_input = squeezed_input_shape[i]
-        if s_input != -1 and s_data != s_input:
+        s_placeholder = squeezed_placeholder_shape[i]
+        if s_placeholder != -1 and s_data != s_placeholder:
             return False
 
     return True

baselines/common/vec_env/vec_env.py

@@ -146,8 +146,8 @@ class VecEnvWrapper(VecEnv):
     def __init__(self, venv, observation_space=None, action_space=None):
         self.venv = venv
         super().__init__(num_envs=venv.num_envs,
-                         observation_space=observation_space or venv.observation_space,
-                         action_space=action_space or venv.action_space)
+                        observation_space=observation_space or venv.observation_space,
+                        action_space=action_space or venv.action_space)
 
     def step_async(self, actions):
         self.venv.step_async(actions)
@@ -169,6 +169,11 @@ class VecEnvWrapper(VecEnv):
     def get_images(self):
         return self.venv.get_images()
 
+    def __getattr__(self, name):
+        if name.startswith('_'):
+            raise AttributeError("attempted to get missing private attribute '{}'".format(name))
+        return getattr(self.venv, name)
+
 class VecEnvObservationWrapper(VecEnvWrapper):
     @abstractmethod
     def process(self, obs):

baselines/common/vec_env/vec_monitor.py

@@ -5,16 +5,18 @@ import time
 from collections import deque
 
 class VecMonitor(VecEnvWrapper):
-    def __init__(self, venv, filename=None, keep_buf=0):
+    def __init__(self, venv, filename=None, keep_buf=0, info_keywords=()):
         VecEnvWrapper.__init__(self, venv)
         self.eprets = None
         self.eplens = None
         self.epcount = 0
         self.tstart = time.time()
         if filename:
-            self.results_writer = ResultsWriter(filename, header={'t_start': self.tstart})
+            self.results_writer = ResultsWriter(filename, header={'t_start': self.tstart},
+                extra_keys=info_keywords)
         else:
             self.results_writer = None
+        self.info_keywords = info_keywords
         self.keep_buf = keep_buf
         if self.keep_buf:
             self.epret_buf = deque([], maxlen=keep_buf)
@@ -30,11 +32,16 @@ class VecMonitor(VecEnvWrapper):
         obs, rews, dones, infos = self.venv.step_wait()
         self.eprets += rews
         self.eplens += 1
-        newinfos = []
-        for (i, (done, ret, eplen, info)) in enumerate(zip(dones, self.eprets, self.eplens, infos)):
-            info = info.copy()
-            if done:
+
+        newinfos = list(infos[:])
+        for i in range(len(dones)):
+            if dones[i]:
+                info = infos[i].copy()
+                ret = self.eprets[i]
+                eplen = self.eplens[i]
                 epinfo = {'r': ret, 'l': eplen, 't': round(time.time() - self.tstart, 6)}
+                for k in self.info_keywords:
+                    epinfo[k] = info[k]
                 info['episode'] = epinfo
                 if self.keep_buf:
                     self.epret_buf.append(ret)
@@ -44,6 +51,5 @@ class VecMonitor(VecEnvWrapper):
                 self.eplens[i] = 0
                 if self.results_writer:
                     self.results_writer.write_row(epinfo)
-            newinfos.append(info)
-
+                newinfos[i] = info
         return obs, rews, dones, newinfos

baselines/common/vec_env/vec_normalize.py

@@ -1,18 +1,22 @@
 from . import VecEnvWrapper
-from baselines.common.running_mean_std import RunningMeanStd
 import numpy as np
 
-
 class VecNormalize(VecEnvWrapper):
     """
     A vectorized wrapper that normalizes the observations
     and returns from an environment.
     """
 
-    def __init__(self, venv, ob=True, ret=True, clipob=10., cliprew=10., gamma=0.99, epsilon=1e-8):
+    def __init__(self, venv, ob=True, ret=True, clipob=10., cliprew=10., gamma=0.99, epsilon=1e-8, use_tf=False):
         VecEnvWrapper.__init__(self, venv)
-        self.ob_rms = RunningMeanStd(shape=self.observation_space.shape) if ob else None
-        self.ret_rms = RunningMeanStd(shape=()) if ret else None
+        if use_tf:
+            from baselines.common.running_mean_std import TfRunningMeanStd
+            self.ob_rms = TfRunningMeanStd(shape=self.observation_space.shape, scope='ob_rms') if ob else None
+            self.ret_rms = TfRunningMeanStd(shape=(), scope='ret_rms') if ret else None
+        else:
+            from baselines.common.running_mean_std import RunningMeanStd
+            self.ob_rms = RunningMeanStd(shape=self.observation_space.shape) if ob else None
+            self.ret_rms = RunningMeanStd(shape=()) if ret else None
         self.clipob = clipob
         self.cliprew = cliprew
         self.ret = np.zeros(self.num_envs)

baselines/common/vec_env/vec_remove_dict_obs.py

@@ -1,6 +1,5 @@
 from .vec_env import VecEnvObservationWrapper
 
-
 class VecExtractDictObs(VecEnvObservationWrapper):
     def __init__(self, venv, key):
         self.key = key
@@ -8,4 +7,4 @@ class VecExtractDictObs(VecEnvObservationWrapper):
             observation_space=venv.observation_space.spaces[self.key])
 
     def process(self, obs):
-        return obs[self.key]
+        return obs[self.key]

baselines/common/wrappers.py

@@ -16,4 +16,14 @@ class TimeLimit(gym.Wrapper):
 
     def reset(self, **kwargs):
         self._elapsed_steps = 0
-        return self.env.reset(**kwargs)
+        return self.env.reset(**kwargs)
+
+class ClipActionsWrapper(gym.Wrapper):
+    def step(self, action):
+        import numpy as np
+        action = np.nan_to_num(action)
+        action = np.clip(action, self.action_space.low, self.action_space.high)
+        return self.env.step(action)
+
+    def reset(self, **kwargs):
+        return self.env.reset(**kwargs)

baselines/ddpg/ddpg.py

@@ -1,5 +1,4 @@
 import os
-import os.path as osp
 import time
 from collections import deque
 import pickle
@@ -9,9 +8,9 @@ from baselines.ddpg.models import Actor, Critic
 from baselines.ddpg.memory import Memory
 from baselines.ddpg.noise import AdaptiveParamNoiseSpec, NormalActionNoise, OrnsteinUhlenbeckActionNoise
 from baselines.common import set_global_seeds
+import baselines.common.tf_util as U
 
 from baselines import logger
-import tensorflow as tf
 import numpy as np
 
 try:
@@ -43,7 +42,6 @@ def learn(network, env,
           tau=0.01,
           eval_env=None,
           param_noise_adaption_interval=50,
-          load_path=None,
           **network_kwargs):
 
     set_global_seeds(seed)
@@ -63,8 +61,8 @@ def learn(network, env,
     assert (np.abs(env.action_space.low) == env.action_space.high).all()  # we assume symmetric actions.
 
     memory = Memory(limit=int(1e6), action_shape=env.action_space.shape, observation_shape=env.observation_space.shape)
-    critic = Critic(nb_actions, ob_shape=env.observation_space.shape, network=network, **network_kwargs)
-    actor = Actor(nb_actions, ob_shape=env.observation_space.shape, network=network, **network_kwargs)
+    critic = Critic(network=network, **network_kwargs)
+    actor = Actor(nb_actions, network=network, **network_kwargs)
 
     action_noise = None
     param_noise = None
@@ -96,18 +94,12 @@ def learn(network, env,
     logger.info('Using agent with the following configuration:')
     logger.info(str(agent.__dict__.items()))
 
-    if load_path is not None:
-        load_path = osp.expanduser(load_path)
-        ckpt = tf.train.Checkpoint(model=agent)
-        manager = tf.train.CheckpointManager(ckpt, load_path, max_to_keep=None)
-        ckpt.restore(manager.latest_checkpoint)
-        print("Restoring from {}".format(manager.latest_checkpoint))
-
     eval_episode_rewards_history = deque(maxlen=100)
     episode_rewards_history = deque(maxlen=100)
-
+    sess = U.get_session()
     # Prepare everything.
-    agent.initialize()
+    agent.initialize(sess)
+    sess.graph.finalize()
 
     agent.reset()
 
@@ -141,8 +133,7 @@ def learn(network, env,
                 agent.reset()
             for t_rollout in range(nb_rollout_steps):
                 # Predict next action.
-                action, q, _, _ = agent.step(tf.constant(obs), apply_noise=True, compute_Q=True)
-                action, q = action.numpy(), q.numpy()
+                action, q, _, _ = agent.step(obs, apply_noise=True, compute_Q=True)
 
                 # Execute next action.
                 if rank == 0 and render:
@@ -179,6 +170,7 @@ def learn(network, env,
                             agent.reset()
 
 
+
             # Train.
             epoch_actor_losses = []
             epoch_critic_losses = []
@@ -186,9 +178,7 @@ def learn(network, env,
             for t_train in range(nb_train_steps):
                 # Adapt param noise, if necessary.
                 if memory.nb_entries >= batch_size and t_train % param_noise_adaption_interval == 0:
-                    batch = agent.memory.sample(batch_size=batch_size)
-                    obs0 = tf.constant(batch['obs0'])
-                    distance = agent.adapt_param_noise(obs0)
+                    distance = agent.adapt_param_noise()
                     epoch_adaptive_distances.append(distance)
 
                 cl, al = agent.train()

baselines/ddpg/ddpg_learner.py

@@ -1,15 +1,16 @@
+from copy import copy
 from functools import reduce
 
 import numpy as np
 import tensorflow as tf
+import tensorflow.contrib as tc
 
 from baselines import logger
-from baselines.ddpg.models import Actor, Critic
+from baselines.common.mpi_adam import MpiAdam
+import baselines.common.tf_util as U
 from baselines.common.mpi_running_mean_std import RunningMeanStd
 try:
     from mpi4py import MPI
-    from baselines.common.mpi_adam_optimizer import MpiAdamOptimizer
-    from baselines.common.mpi_util import sync_from_root
 except ImportError:
     MPI = None
 
@@ -24,7 +25,6 @@ def denormalize(x, stats):
         return x
     return x * stats.std + stats.mean
 
-@tf.function
 def reduce_std(x, axis=None, keepdims=False):
     return tf.sqrt(reduce_var(x, axis=axis, keepdims=keepdims))
 
@@ -33,21 +33,49 @@ def reduce_var(x, axis=None, keepdims=False):
     devs_squared = tf.square(x - m)
     return tf.reduce_mean(devs_squared, axis=axis, keepdims=keepdims)
 
-@tf.function
-def update_perturbed_actor(actor, perturbed_actor, param_noise_stddev):
+def get_target_updates(vars, target_vars, tau):
+    logger.info('setting up target updates ...')
+    soft_updates = []
+    init_updates = []
+    assert len(vars) == len(target_vars)
+    for var, target_var in zip(vars, target_vars):
+        logger.info('  {} <- {}'.format(target_var.name, var.name))
+        init_updates.append(tf.assign(target_var, var))
+        soft_updates.append(tf.assign(target_var, (1. - tau) * target_var + tau * var))
+    assert len(init_updates) == len(vars)
+    assert len(soft_updates) == len(vars)
+    return tf.group(*init_updates), tf.group(*soft_updates)
 
-    for var, perturbed_var in zip(actor.variables, perturbed_actor.variables):
+
+def get_perturbed_actor_updates(actor, perturbed_actor, param_noise_stddev):
+    assert len(actor.vars) == len(perturbed_actor.vars)
+    assert len(actor.perturbable_vars) == len(perturbed_actor.perturbable_vars)
+
+    updates = []
+    for var, perturbed_var in zip(actor.vars, perturbed_actor.vars):
         if var in actor.perturbable_vars:
-            perturbed_var.assign(var + tf.random.normal(shape=tf.shape(var), mean=0., stddev=param_noise_stddev))
+            logger.info('  {} <- {} + noise'.format(perturbed_var.name, var.name))
+            updates.append(tf.assign(perturbed_var, var + tf.random_normal(tf.shape(var), mean=0., stddev=param_noise_stddev)))
         else:
-            perturbed_var.assign(var)
+            logger.info('  {} <- {}'.format(perturbed_var.name, var.name))
+            updates.append(tf.assign(perturbed_var, var))
+    assert len(updates) == len(actor.vars)
+    return tf.group(*updates)
 
 
-class DDPG(tf.Module):
+class DDPG(object):
     def __init__(self, actor, critic, memory, observation_shape, action_shape, param_noise=None, action_noise=None,
         gamma=0.99, tau=0.001, normalize_returns=False, enable_popart=False, normalize_observations=True,
         batch_size=128, observation_range=(-5., 5.), action_range=(-1., 1.), return_range=(-np.inf, np.inf),
         critic_l2_reg=0., actor_lr=1e-4, critic_lr=1e-3, clip_norm=None, reward_scale=1.):
+        # Inputs.
+        self.obs0 = tf.placeholder(tf.float32, shape=(None,) + observation_shape, name='obs0')
+        self.obs1 = tf.placeholder(tf.float32, shape=(None,) + observation_shape, name='obs1')
+        self.terminals1 = tf.placeholder(tf.float32, shape=(None, 1), name='terminals1')
+        self.rewards = tf.placeholder(tf.float32, shape=(None, 1), name='rewards')
+        self.actions = tf.placeholder(tf.float32, shape=(None,) + action_shape, name='actions')
+        self.critic_target = tf.placeholder(tf.float32, shape=(None, 1), name='critic_target')
+        self.param_noise_stddev = tf.placeholder(tf.float32, shape=(), name='param_noise_stddev')
 
         # Parameters.
         self.gamma = gamma
@@ -60,103 +88,128 @@ class DDPG(tf.Module):
         self.action_range = action_range
         self.return_range = return_range
         self.observation_range = observation_range
-        self.observation_shape = observation_shape
         self.critic = critic
         self.actor = actor
+        self.actor_lr = actor_lr
+        self.critic_lr = critic_lr
         self.clip_norm = clip_norm
         self.enable_popart = enable_popart
         self.reward_scale = reward_scale
         self.batch_size = batch_size
         self.stats_sample = None
         self.critic_l2_reg = critic_l2_reg
-        self.actor_lr = tf.constant(actor_lr)
-        self.critic_lr = tf.constant(critic_lr)
 
         # Observation normalization.
         if self.normalize_observations:
-            with tf.name_scope('obs_rms'):
+            with tf.variable_scope('obs_rms'):
                 self.obs_rms = RunningMeanStd(shape=observation_shape)
         else:
             self.obs_rms = None
+        normalized_obs0 = tf.clip_by_value(normalize(self.obs0, self.obs_rms),
+            self.observation_range[0], self.observation_range[1])
+        normalized_obs1 = tf.clip_by_value(normalize(self.obs1, self.obs_rms),
+            self.observation_range[0], self.observation_range[1])
 
         # Return normalization.
         if self.normalize_returns:
-            with tf.name_scope('ret_rms'):
+            with tf.variable_scope('ret_rms'):
                 self.ret_rms = RunningMeanStd()
         else:
             self.ret_rms = None
 
         # Create target networks.
-        self.target_critic = Critic(actor.nb_actions, observation_shape, name='target_critic', network=critic.network, **critic.network_kwargs)
-        self.target_actor = Actor(actor.nb_actions, observation_shape, name='target_actor', network=actor.network, **actor.network_kwargs)
+        target_actor = copy(actor)
+        target_actor.name = 'target_actor'
+        self.target_actor = target_actor
+        target_critic = copy(critic)
+        target_critic.name = 'target_critic'
+        self.target_critic = target_critic
+
+        # Create networks and core TF parts that are shared across setup parts.
+        self.actor_tf = actor(normalized_obs0)
+        self.normalized_critic_tf = critic(normalized_obs0, self.actions)
+        self.critic_tf = denormalize(tf.clip_by_value(self.normalized_critic_tf, self.return_range[0], self.return_range[1]), self.ret_rms)
+        self.normalized_critic_with_actor_tf = critic(normalized_obs0, self.actor_tf, reuse=True)
+        self.critic_with_actor_tf = denormalize(tf.clip_by_value(self.normalized_critic_with_actor_tf, self.return_range[0], self.return_range[1]), self.ret_rms)
+        Q_obs1 = denormalize(target_critic(normalized_obs1, target_actor(normalized_obs1)), self.ret_rms)
+        self.target_Q = self.rewards + (1. - self.terminals1) * gamma * Q_obs1
 
         # Set up parts.
         if self.param_noise is not None:
-            self.setup_param_noise()
-
-        if MPI is not None:
-            comm = MPI.COMM_WORLD
-            self.actor_optimizer = MpiAdamOptimizer(comm, self.actor.trainable_variables)
-            self.critic_optimizer = MpiAdamOptimizer(comm, self.critic.trainable_variables)
-        else:
-            self.actor_optimizer = tf.keras.optimizers.Adam(learning_rate=actor_lr)
-            self.critic_optimizer = tf.keras.optimizers.Adam(learning_rate=critic_lr)
-
-        logger.info('setting up actor optimizer')
-        actor_shapes = [var.get_shape().as_list() for var in self.actor.trainable_variables]
-        actor_nb_params = sum([reduce(lambda x, y: x * y, shape) for shape in actor_shapes])
-        logger.info('  actor shapes: {}'.format(actor_shapes))
-        logger.info('  actor params: {}'.format(actor_nb_params))
-        logger.info('setting up critic optimizer')
-        critic_shapes = [var.get_shape().as_list() for var in self.critic.trainable_variables]
-        critic_nb_params = sum([reduce(lambda x, y: x * y, shape) for shape in critic_shapes])
-        logger.info('  critic shapes: {}'.format(critic_shapes))
-        logger.info('  critic params: {}'.format(critic_nb_params))
-        if self.critic_l2_reg > 0.:
-            critic_reg_vars = []
-            for layer in self.critic.network_builder.layers[1:]:
-                critic_reg_vars.append(layer.kernel)
-            for var in critic_reg_vars:
-                logger.info('  regularizing: {}'.format(var.name))
-            logger.info('  applying l2 regularization with {}'.format(self.critic_l2_reg))
-
-        logger.info('setting up critic target updates ...')
-        for var, target_var in zip(self.critic.variables, self.target_critic.variables):
-            logger.info('  {} <- {}'.format(target_var.name, var.name))
-        logger.info('setting up actor target updates ...')
-        for var, target_var in zip(self.actor.variables, self.target_actor.variables):
-            logger.info('  {} <- {}'.format(target_var.name, var.name))
-
-        if self.param_noise:
-            logger.info('setting up param noise')
-            for var, perturbed_var in zip(self.actor.variables, self.perturbed_actor.variables):
-                if var in actor.perturbable_vars:
-                    logger.info('  {} <- {} + noise'.format(perturbed_var.name, var.name))
-                else:
-                    logger.info('  {} <- {}'.format(perturbed_var.name, var.name))
-            for var, perturbed_var in zip(self.actor.variables, self.perturbed_adaptive_actor.variables):
-                if var in actor.perturbable_vars:
-                    logger.info('  {} <- {} + noise'.format(perturbed_var.name, var.name))
-                else:
-                    logger.info('  {} <- {}'.format(perturbed_var.name, var.name))
-
+            self.setup_param_noise(normalized_obs0)
+        self.setup_actor_optimizer()
+        self.setup_critic_optimizer()
         if self.normalize_returns and self.enable_popart:
             self.setup_popart()
+        self.setup_stats()
+        self.setup_target_network_updates()
 
         self.initial_state = None # recurrent architectures not supported yet
 
+    def setup_target_network_updates(self):
+        actor_init_updates, actor_soft_updates = get_target_updates(self.actor.vars, self.target_actor.vars, self.tau)
+        critic_init_updates, critic_soft_updates = get_target_updates(self.critic.vars, self.target_critic.vars, self.tau)
+        self.target_init_updates = [actor_init_updates, critic_init_updates]
+        self.target_soft_updates = [actor_soft_updates, critic_soft_updates]
 
-    def setup_param_noise(self):
+    def setup_param_noise(self, normalized_obs0):
         assert self.param_noise is not None
 
         # Configure perturbed actor.
-        self.perturbed_actor = Actor(self.actor.nb_actions, self.observation_shape, name='param_noise_actor', network=self.actor.network, **self.actor.network_kwargs)
+        param_noise_actor = copy(self.actor)
+        param_noise_actor.name = 'param_noise_actor'
+        self.perturbed_actor_tf = param_noise_actor(normalized_obs0)
+        logger.info('setting up param noise')
+        self.perturb_policy_ops = get_perturbed_actor_updates(self.actor, param_noise_actor, self.param_noise_stddev)
 
         # Configure separate copy for stddev adoption.
-        self.perturbed_adaptive_actor = Actor(self.actor.nb_actions, self.observation_shape, name='adaptive_param_noise_actor', network=self.actor.network, **self.actor.network_kwargs)
+        adaptive_param_noise_actor = copy(self.actor)
+        adaptive_param_noise_actor.name = 'adaptive_param_noise_actor'
+        adaptive_actor_tf = adaptive_param_noise_actor(normalized_obs0)
+        self.perturb_adaptive_policy_ops = get_perturbed_actor_updates(self.actor, adaptive_param_noise_actor, self.param_noise_stddev)
+        self.adaptive_policy_distance = tf.sqrt(tf.reduce_mean(tf.square(self.actor_tf - adaptive_actor_tf)))
+
+    def setup_actor_optimizer(self):
+        logger.info('setting up actor optimizer')
+        self.actor_loss = -tf.reduce_mean(self.critic_with_actor_tf)
+        actor_shapes = [var.get_shape().as_list() for var in self.actor.trainable_vars]
+        actor_nb_params = sum([reduce(lambda x, y: x * y, shape) for shape in actor_shapes])
+        logger.info('  actor shapes: {}'.format(actor_shapes))
+        logger.info('  actor params: {}'.format(actor_nb_params))
+        self.actor_grads = U.flatgrad(self.actor_loss, self.actor.trainable_vars, clip_norm=self.clip_norm)
+        self.actor_optimizer = MpiAdam(var_list=self.actor.trainable_vars,
+            beta1=0.9, beta2=0.999, epsilon=1e-08)
+
+    def setup_critic_optimizer(self):
+        logger.info('setting up critic optimizer')
+        normalized_critic_target_tf = tf.clip_by_value(normalize(self.critic_target, self.ret_rms), self.return_range[0], self.return_range[1])
+        self.critic_loss = tf.reduce_mean(tf.square(self.normalized_critic_tf - normalized_critic_target_tf))
+        if self.critic_l2_reg > 0.:
+            critic_reg_vars = [var for var in self.critic.trainable_vars if var.name.endswith('/w:0') and 'output' not in var.name]
+            for var in critic_reg_vars:
+                logger.info('  regularizing: {}'.format(var.name))
+            logger.info('  applying l2 regularization with {}'.format(self.critic_l2_reg))
+            critic_reg = tc.layers.apply_regularization(
+                tc.layers.l2_regularizer(self.critic_l2_reg),
+                weights_list=critic_reg_vars
+            )
+            self.critic_loss += critic_reg
+        critic_shapes = [var.get_shape().as_list() for var in self.critic.trainable_vars]
+        critic_nb_params = sum([reduce(lambda x, y: x * y, shape) for shape in critic_shapes])
+        logger.info('  critic shapes: {}'.format(critic_shapes))
+        logger.info('  critic params: {}'.format(critic_nb_params))
+        self.critic_grads = U.flatgrad(self.critic_loss, self.critic.trainable_vars, clip_norm=self.clip_norm)
+        self.critic_optimizer = MpiAdam(var_list=self.critic.trainable_vars,
+            beta1=0.9, beta2=0.999, epsilon=1e-08)
 
     def setup_popart(self):
         # See https://arxiv.org/pdf/1602.07714.pdf for details.
+        self.old_std = tf.placeholder(tf.float32, shape=[1], name='old_std')
+        new_std = self.ret_rms.std
+        self.old_mean = tf.placeholder(tf.float32, shape=[1], name='old_mean')
+        new_mean = self.ret_rms.mean
+
+        self.renormalize_Q_outputs_op = []
         for vs in [self.critic.output_vars, self.target_critic.output_vars]:
             assert len(vs) == 2
             M, b = vs
@@ -164,26 +217,63 @@ class DDPG(tf.Module):
             assert 'bias' in b.name
             assert M.get_shape()[-1] == 1
             assert b.get_shape()[-1] == 1
+            self.renormalize_Q_outputs_op += [M.assign(M * self.old_std / new_std)]
+            self.renormalize_Q_outputs_op += [b.assign((b * self.old_std + self.old_mean - new_mean) / new_std)]
+
+    def setup_stats(self):
+        ops = []
+        names = []
+
+        if self.normalize_returns:
+            ops += [self.ret_rms.mean, self.ret_rms.std]
+            names += ['ret_rms_mean', 'ret_rms_std']
+
+        if self.normalize_observations:
+            ops += [tf.reduce_mean(self.obs_rms.mean), tf.reduce_mean(self.obs_rms.std)]
+            names += ['obs_rms_mean', 'obs_rms_std']
+
+        ops += [tf.reduce_mean(self.critic_tf)]
+        names += ['reference_Q_mean']
+        ops += [reduce_std(self.critic_tf)]
+        names += ['reference_Q_std']
+
+        ops += [tf.reduce_mean(self.critic_with_actor_tf)]
+        names += ['reference_actor_Q_mean']
+        ops += [reduce_std(self.critic_with_actor_tf)]
+        names += ['reference_actor_Q_std']
+
+        ops += [tf.reduce_mean(self.actor_tf)]
+        names += ['reference_action_mean']
+        ops += [reduce_std(self.actor_tf)]
+        names += ['reference_action_std']
+
+        if self.param_noise:
+            ops += [tf.reduce_mean(self.perturbed_actor_tf)]
+            names += ['reference_perturbed_action_mean']
+            ops += [reduce_std(self.perturbed_actor_tf)]
+            names += ['reference_perturbed_action_std']
+
+        self.stats_ops = ops
+        self.stats_names = names
 
-    @tf.function
     def step(self, obs, apply_noise=True, compute_Q=True):
-        normalized_obs = tf.clip_by_value(normalize(obs, self.obs_rms), self.observation_range[0], self.observation_range[1])
-        actor_tf = self.actor(normalized_obs)
         if self.param_noise is not None and apply_noise:
-            action = self.perturbed_actor(normalized_obs)
+            actor_tf = self.perturbed_actor_tf
         else:
-            action = actor_tf
-
+            actor_tf = self.actor_tf
+        feed_dict = {self.obs0: U.adjust_shape(self.obs0, [obs])}
         if compute_Q:
-            normalized_critic_with_actor_tf = self.critic(normalized_obs, actor_tf)
-            q = denormalize(tf.clip_by_value(normalized_critic_with_actor_tf, self.return_range[0], self.return_range[1]), self.ret_rms)
+            action, q = self.sess.run([actor_tf, self.critic_with_actor_tf], feed_dict=feed_dict)
         else:
+            action = self.sess.run(actor_tf, feed_dict=feed_dict)
             q = None
 
         if self.action_noise is not None and apply_noise:
             noise = self.action_noise()
+            assert noise.shape == action[0].shape
             action += noise
-        action = tf.clip_by_value(action, self.action_range[0], self.action_range[1])
+        action = np.clip(action, self.action_range[0], self.action_range[1])
+
 
         return action, q, None, None
 
@@ -197,130 +287,79 @@ class DDPG(tf.Module):
                 self.obs_rms.update(np.array([obs0[b]]))
 
     def train(self):
+        # Get a batch.
         batch = self.memory.sample(batch_size=self.batch_size)
-        obs0, obs1 = tf.constant(batch['obs0']), tf.constant(batch['obs1'])
-        actions, rewards, terminals1 = tf.constant(batch['actions']), tf.constant(batch['rewards']), tf.constant(batch['terminals1'], dtype=tf.float32)
-        normalized_obs0, target_Q = self.compute_normalized_obs0_and_target_Q(obs0, obs1, rewards, terminals1)
 
         if self.normalize_returns and self.enable_popart:
-            old_mean = self.ret_rms.mean
-            old_std = self.ret_rms.std
+            old_mean, old_std, target_Q = self.sess.run([self.ret_rms.mean, self.ret_rms.std, self.target_Q], feed_dict={
+                self.obs1: batch['obs1'],
+                self.rewards: batch['rewards'],
+                self.terminals1: batch['terminals1'].astype('float32'),
+            })
             self.ret_rms.update(target_Q.flatten())
-            # renormalize Q outputs
-            new_mean = self.ret_rms.mean
-            new_std = self.ret_rms.std
-            for vs in [self.critic.output_vars, self.target_critic.output_vars]:
-                kernel, bias = vs
-                kernel.assign(kernel * old_std / new_std)
-                bias.assign((bias * old_std + old_mean - new_mean) / new_std)
+            self.sess.run(self.renormalize_Q_outputs_op, feed_dict={
+                self.old_std : np.array([old_std]),
+                self.old_mean : np.array([old_mean]),
+            })
 
-
-        actor_grads, actor_loss = self.get_actor_grads(normalized_obs0)
-        critic_grads, critic_loss = self.get_critic_grads(normalized_obs0, actions, target_Q)
-
-        if MPI is not None:
-            self.actor_optimizer.apply_gradients(actor_grads, self.actor_lr)
-            self.critic_optimizer.apply_gradients(critic_grads, self.critic_lr)
+            # Run sanity check. Disabled by default since it slows down things considerably.
+            # print('running sanity check')
+            # target_Q_new, new_mean, new_std = self.sess.run([self.target_Q, self.ret_rms.mean, self.ret_rms.std], feed_dict={
+            #     self.obs1: batch['obs1'],
+            #     self.rewards: batch['rewards'],
+            #     self.terminals1: batch['terminals1'].astype('float32'),
+            # })
+            # print(target_Q_new, target_Q, new_mean, new_std)
+            # assert (np.abs(target_Q - target_Q_new) < 1e-3).all()
         else:
-            self.actor_optimizer.apply_gradients(zip(actor_grads, self.actor.trainable_variables))
-            self.critic_optimizer.apply_gradients(zip(critic_grads, self.critic.trainable_variables))
+            target_Q = self.sess.run(self.target_Q, feed_dict={
+                self.obs1: batch['obs1'],
+                self.rewards: batch['rewards'],
+                self.terminals1: batch['terminals1'].astype('float32'),
+            })
+
+        # Get all gradients and perform a synced update.
+        ops = [self.actor_grads, self.actor_loss, self.critic_grads, self.critic_loss]
+        actor_grads, actor_loss, critic_grads, critic_loss = self.sess.run(ops, feed_dict={
+            self.obs0: batch['obs0'],
+            self.actions: batch['actions'],
+            self.critic_target: target_Q,
+        })
+        self.actor_optimizer.update(actor_grads, stepsize=self.actor_lr)
+        self.critic_optimizer.update(critic_grads, stepsize=self.critic_lr)
 
         return critic_loss, actor_loss
 
-    @tf.function
-    def compute_normalized_obs0_and_target_Q(self, obs0, obs1, rewards, terminals1):
-        normalized_obs0 = tf.clip_by_value(normalize(obs0, self.obs_rms), self.observation_range[0], self.observation_range[1])
-        normalized_obs1 = tf.clip_by_value(normalize(obs1, self.obs_rms), self.observation_range[0], self.observation_range[1])
-        Q_obs1 = denormalize(self.target_critic(normalized_obs1, self.target_actor(normalized_obs1)), self.ret_rms)
-        target_Q = rewards + (1. - terminals1) * self.gamma * Q_obs1
-        return normalized_obs0, target_Q
+    def initialize(self, sess):
+        self.sess = sess
+        self.sess.run(tf.global_variables_initializer())
+        self.actor_optimizer.sync()
+        self.critic_optimizer.sync()
+        self.sess.run(self.target_init_updates)
 
-    @tf.function
-    def get_actor_grads(self, normalized_obs0):
-        with tf.GradientTape() as tape:
-            actor_tf = self.actor(normalized_obs0)
-            normalized_critic_with_actor_tf = self.critic(normalized_obs0, actor_tf)
-            critic_with_actor_tf = denormalize(tf.clip_by_value(normalized_critic_with_actor_tf, self.return_range[0], self.return_range[1]), self.ret_rms)
-            actor_loss = -tf.reduce_mean(critic_with_actor_tf)
-        actor_grads = tape.gradient(actor_loss, self.actor.trainable_variables)
-        if self.clip_norm:
-            actor_grads = [tf.clip_by_norm(grad, clip_norm=self.clip_norm) for grad in actor_grads]
-        if MPI is not None:
-            actor_grads = tf.concat([tf.reshape(g, (-1,)) for g in actor_grads], axis=0)
-        return actor_grads, actor_loss
-
-    @tf.function
-    def get_critic_grads(self, normalized_obs0, actions, target_Q):
-        with tf.GradientTape() as tape:
-            normalized_critic_tf = self.critic(normalized_obs0, actions)
-            normalized_critic_target_tf = tf.clip_by_value(normalize(target_Q, self.ret_rms), self.return_range[0], self.return_range[1])
-            critic_loss = tf.reduce_mean(tf.square(normalized_critic_tf - normalized_critic_target_tf))
-            # The first is input layer, which is ignored here.
-            if self.critic_l2_reg > 0.:
-                # Ignore the first input layer.
-                for layer in self.critic.network_builder.layers[1:]:
-                    # The original l2_regularizer takes half of sum square.
-                    critic_loss += (self.critic_l2_reg / 2.)* tf.reduce_sum(tf.square(layer.kernel))
-        critic_grads = tape.gradient(critic_loss, self.critic.trainable_variables)
-        if self.clip_norm:
-            critic_grads = [tf.clip_by_norm(grad, clip_norm=self.clip_norm) for grad in critic_grads]
-        if MPI is not None:
-            critic_grads = tf.concat([tf.reshape(g, (-1,)) for g in critic_grads], axis=0)
-        return critic_grads, critic_loss
-
-
-    def initialize(self):
-        if MPI is not None:
-            sync_from_root(self.actor.trainable_variables + self.critic.trainable_variables)
-        self.target_actor.set_weights(self.actor.get_weights())
-        self.target_critic.set_weights(self.critic.get_weights())
-
-    @tf.function
     def update_target_net(self):
-        for var, target_var in zip(self.actor.variables, self.target_actor.variables):
-            target_var.assign((1. - self.tau) * target_var + self.tau * var)
-        for var, target_var in zip(self.critic.variables, self.target_critic.variables):
-            target_var.assign((1. - self.tau) * target_var + self.tau * var)
+        self.sess.run(self.target_soft_updates)
 
     def get_stats(self):
-
         if self.stats_sample is None:
             # Get a sample and keep that fixed for all further computations.
             # This allows us to estimate the change in value for the same set of inputs.
             self.stats_sample = self.memory.sample(batch_size=self.batch_size)
-        obs0 = self.stats_sample['obs0']
-        actions = self.stats_sample['actions']
-        normalized_obs0 = tf.clip_by_value(normalize(obs0, self.obs_rms), self.observation_range[0], self.observation_range[1])
-        normalized_critic_tf = self.critic(normalized_obs0, actions)
-        critic_tf = denormalize(tf.clip_by_value(normalized_critic_tf, self.return_range[0], self.return_range[1]), self.ret_rms)
-        actor_tf = self.actor(normalized_obs0)
-        normalized_critic_with_actor_tf = self.critic(normalized_obs0, actor_tf)
-        critic_with_actor_tf = denormalize(tf.clip_by_value(normalized_critic_with_actor_tf, self.return_range[0], self.return_range[1]), self.ret_rms)
+        values = self.sess.run(self.stats_ops, feed_dict={
+            self.obs0: self.stats_sample['obs0'],
+            self.actions: self.stats_sample['actions'],
+        })
 
-        stats = {}
-        if self.normalize_returns:
-            stats['ret_rms_mean'] = self.ret_rms.mean
-            stats['ret_rms_std'] = self.ret_rms.std
-        if self.normalize_observations:
-            stats['obs_rms_mean'] = tf.reduce_mean(self.obs_rms.mean)
-            stats['obs_rms_std'] = tf.reduce_mean(self.obs_rms.std)
-        stats['reference_Q_mean'] = tf.reduce_mean(critic_tf)
-        stats['reference_Q_std'] = reduce_std(critic_tf)
-        stats['reference_actor_Q_mean'] = tf.reduce_mean(critic_with_actor_tf)
-        stats['reference_actor_Q_std'] = reduce_std(critic_with_actor_tf)
-        stats['reference_action_mean'] = tf.reduce_mean(actor_tf)
-        stats['reference_action_std'] = reduce_std(actor_tf)
+        names = self.stats_names[:]
+        assert len(names) == len(values)
+        stats = dict(zip(names, values))
+
+        if self.param_noise is not None:
+            stats = {**stats, **self.param_noise.get_stats()}
 
-        if self.param_noise:
-            perturbed_actor_tf = self.perturbed_actor(normalized_obs0)
-            stats['reference_perturbed_action_mean'] = tf.reduce_mean(perturbed_actor_tf)
-            stats['reference_perturbed_action_std'] = reduce_std(perturbed_actor_tf)
-            stats.update(self.param_noise.get_stats())
         return stats
 
-
-
-    def adapt_param_noise(self, obs0):
+    def adapt_param_noise(self):
         try:
             from mpi4py import MPI
         except ImportError:
@@ -329,28 +368,34 @@ class DDPG(tf.Module):
         if self.param_noise is None:
             return 0.
 
-        mean_distance = self.get_mean_distance(obs0).numpy()
+        # Perturb a separate copy of the policy to adjust the scale for the next "real" perturbation.
+        batch = self.memory.sample(batch_size=self.batch_size)
+        self.sess.run(self.perturb_adaptive_policy_ops, feed_dict={
+            self.param_noise_stddev: self.param_noise.current_stddev,
+        })
+        distance = self.sess.run(self.adaptive_policy_distance, feed_dict={
+            self.obs0: batch['obs0'],
+            self.param_noise_stddev: self.param_noise.current_stddev,
+        })
 
         if MPI is not None:
-            mean_distance = MPI.COMM_WORLD.allreduce(mean_distance, op=MPI.SUM) / MPI.COMM_WORLD.Get_size()
+            mean_distance = MPI.COMM_WORLD.allreduce(distance, op=MPI.SUM) / MPI.COMM_WORLD.Get_size()
+        else:
+            mean_distance = distance
+
+        if MPI is not None:
+            mean_distance = MPI.COMM_WORLD.allreduce(distance, op=MPI.SUM) / MPI.COMM_WORLD.Get_size()
+        else:
+            mean_distance = distance
 
         self.param_noise.adapt(mean_distance)
         return mean_distance
 
-    @tf.function
-    def get_mean_distance(self, obs0):
-        # Perturb a separate copy of the policy to adjust the scale for the next "real" perturbation.
-        update_perturbed_actor(self.actor, self.perturbed_adaptive_actor, self.param_noise.current_stddev)
-
-        normalized_obs0 = tf.clip_by_value(normalize(obs0, self.obs_rms), self.observation_range[0], self.observation_range[1])
-        actor_tf = self.actor(normalized_obs0)
-        adaptive_actor_tf = self.perturbed_adaptive_actor(normalized_obs0)
-        mean_distance = tf.sqrt(tf.reduce_mean(tf.square(actor_tf - adaptive_actor_tf)))
-        return mean_distance
-
     def reset(self):
         # Reset internal state after an episode is complete.
         if self.action_noise is not None:
             self.action_noise.reset()
         if self.param_noise is not None:
-            update_perturbed_actor(self.actor, self.perturbed_actor, self.param_noise.current_stddev)
+            self.sess.run(self.perturb_policy_ops, feed_dict={
+                self.param_noise_stddev: self.param_noise.current_stddev,
+            })

baselines/ddpg/models.py

@@ -2,48 +2,50 @@ import tensorflow as tf
 from baselines.common.models import get_network_builder
 
 
-class Model(tf.keras.Model):
+class Model(object):
     def __init__(self, name, network='mlp', **network_kwargs):
-        super(Model, self).__init__(name=name)
-        self.network = network
-        self.network_kwargs = network_kwargs
+        self.name = name
+        self.network_builder = get_network_builder(network)(**network_kwargs)
+
+    @property
+    def vars(self):
+        return tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope=self.name)
+
+    @property
+    def trainable_vars(self):
+        return tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope=self.name)
 
     @property
     def perturbable_vars(self):
-        return [var for var in self.trainable_variables if 'layer_normalization' not in var.name]
+        return [var for var in self.trainable_vars if 'LayerNorm' not in var.name]
 
 
 class Actor(Model):
-    def __init__(self, nb_actions, ob_shape, name='actor', network='mlp', **network_kwargs):
+    def __init__(self, nb_actions, name='actor', network='mlp', **network_kwargs):
         super().__init__(name=name, network=network, **network_kwargs)
         self.nb_actions = nb_actions
-        self.network_builder = get_network_builder(network)(**network_kwargs)(ob_shape)
-        self.output_layer = tf.keras.layers.Dense(units=self.nb_actions,
-                                                  activation=tf.keras.activations.tanh,
-                                                  kernel_initializer=tf.random_uniform_initializer(minval=-3e-3, maxval=3e-3))
-        _ = self.output_layer(self.network_builder.outputs[0])
 
-    @tf.function
-    def call(self, obs):
-        return self.output_layer(self.network_builder(obs))
+    def __call__(self, obs, reuse=False):
+        with tf.variable_scope(self.name, reuse=tf.AUTO_REUSE):
+            x = self.network_builder(obs)
+            x = tf.layers.dense(x, self.nb_actions, kernel_initializer=tf.random_uniform_initializer(minval=-3e-3, maxval=3e-3))
+            x = tf.nn.tanh(x)
+        return x
 
 
 class Critic(Model):
-    def __init__(self, nb_actions, ob_shape, name='critic', network='mlp', **network_kwargs):
+    def __init__(self, name='critic', network='mlp', **network_kwargs):
         super().__init__(name=name, network=network, **network_kwargs)
         self.layer_norm = True
-        self.network_builder = get_network_builder(network)(**network_kwargs)((ob_shape[0] + nb_actions,))
-        self.output_layer = tf.keras.layers.Dense(units=1,
-                                                  kernel_initializer=tf.random_uniform_initializer(minval=-3e-3, maxval=3e-3),
-                                                  name='output')
-        _ = self.output_layer(self.network_builder.outputs[0])
 
-    @tf.function
-    def call(self, obs, actions):
-        x = tf.concat([obs, actions], axis=-1) # this assumes observation and action can be concatenated
-        x = self.network_builder(x)
-        return self.output_layer(x)
+    def __call__(self, obs, action, reuse=False):
+        with tf.variable_scope(self.name, reuse=tf.AUTO_REUSE):
+            x = tf.concat([obs, action], axis=-1) # this assumes observation and action can be concatenated
+            x = self.network_builder(x)
+            x = tf.layers.dense(x, 1, kernel_initializer=tf.random_uniform_initializer(minval=-3e-3, maxval=3e-3), name='output')
+        return x
 
     @property
     def output_vars(self):
-        return self.output_layer.trainable_variables
+        output_vars = [var for var in self.trainable_vars if 'output' in var.name]
+        return output_vars

baselines/ddpg/test_smoke.py

@@ -0,0 +1,16 @@
+from baselines.common.tests.util import smoketest
+def _run(argstr):
+    smoketest('--alg=ddpg --env=Pendulum-v0 --num_timesteps=0 ' + argstr)
+
+def test_popart():
+    _run('--normalize_returns=True --popart=True')
+
+def test_noise_normal():
+    _run('--noise_type=normal_0.1')
+
+def test_noise_ou():
+    _run('--noise_type=ou_0.1')
+
+def test_noise_adaptive():
+    _run('--noise_type=adaptive-param_0.2,normal_0.1')
+

baselines/deepq/__init__.py

@@ -1,7 +1,7 @@
-from baselines.deepq import models  # noqa F401
-from baselines.deepq.deepq_learner import DEEPQ  # noqa F401
-from baselines.deepq.deepq import learn  # noqa F401
-from baselines.deepq.replay_buffer import ReplayBuffer, PrioritizedReplayBuffer  # noqa F401
+from baselines.deepq import models  # noqa
+from baselines.deepq.build_graph import build_act, build_train  # noqa
+from baselines.deepq.deepq import learn, load_act  # noqa
+from baselines.deepq.replay_buffer import ReplayBuffer, PrioritizedReplayBuffer  # noqa
 
 def wrap_atari_dqn(env):
     from baselines.common.atari_wrappers import wrap_deepmind

baselines/deepq/build_graph.py

@@ -0,0 +1,449 @@
+"""Deep Q learning graph
+
+The functions in this file can are used to create the following functions:
+
+======= act ========
+
+    Function to chose an action given an observation
+
+    Parameters
+    ----------
+    observation: object
+        Observation that can be feed into the output of make_obs_ph
+    stochastic: bool
+        if set to False all the actions are always deterministic (default False)
+    update_eps_ph: float
+        update epsilon a new value, if negative not update happens
+        (default: no update)
+
+    Returns
+    -------
+    Tensor of dtype tf.int64 and shape (BATCH_SIZE,) with an action to be performed for
+    every element of the batch.
+
+
+======= act (in case of parameter noise) ========
+
+    Function to chose an action given an observation
+
+    Parameters
+    ----------
+    observation: object
+        Observation that can be feed into the output of make_obs_ph
+    stochastic: bool
+        if set to False all the actions are always deterministic (default False)
+    update_eps_ph: float
+        update epsilon to a new value, if negative no update happens
+        (default: no update)
+    reset_ph: bool
+        reset the perturbed policy by sampling a new perturbation
+    update_param_noise_threshold_ph: float
+        the desired threshold for the difference between non-perturbed and perturbed policy
+    update_param_noise_scale_ph: bool
+        whether or not to update the scale of the noise for the next time it is re-perturbed
+
+    Returns
+    -------
+    Tensor of dtype tf.int64 and shape (BATCH_SIZE,) with an action to be performed for
+    every element of the batch.
+
+
+======= train =======
+
+    Function that takes a transition (s,a,r,s') and optimizes Bellman equation's error:
+
+        td_error = Q(s,a) - (r + gamma * max_a' Q(s', a'))
+        loss = huber_loss[td_error]
+
+    Parameters
+    ----------
+    obs_t: object
+        a batch of observations
+    action: np.array
+        actions that were selected upon seeing obs_t.
+        dtype must be int32 and shape must be (batch_size,)
+    reward: np.array
+        immediate reward attained after executing those actions
+        dtype must be float32 and shape must be (batch_size,)
+    obs_tp1: object
+        observations that followed obs_t
+    done: np.array
+        1 if obs_t was the last observation in the episode and 0 otherwise
+        obs_tp1 gets ignored, but must be of the valid shape.
+        dtype must be float32 and shape must be (batch_size,)
+    weight: np.array
+        imporance weights for every element of the batch (gradient is multiplied
+        by the importance weight) dtype must be float32 and shape must be (batch_size,)
+
+    Returns
+    -------
+    td_error: np.array
+        a list of differences between Q(s,a) and the target in Bellman's equation.
+        dtype is float32 and shape is (batch_size,)
+
+======= update_target ========
+
+    copy the parameters from optimized Q function to the target Q function.
+    In Q learning we actually optimize the following error:
+
+        Q(s,a) - (r + gamma * max_a' Q'(s', a'))
+
+    Where Q' is lagging behind Q to stablize the learning. For example for Atari
+
+    Q' is set to Q once every 10000 updates training steps.
+
+"""
+import tensorflow as tf
+import baselines.common.tf_util as U
+
+
+def scope_vars(scope, trainable_only=False):
+    """
+    Get variables inside a scope
+    The scope can be specified as a string
+    Parameters
+    ----------
+    scope: str or VariableScope
+        scope in which the variables reside.
+    trainable_only: bool
+        whether or not to return only the variables that were marked as trainable.
+    Returns
+    -------
+    vars: [tf.Variable]
+        list of variables in `scope`.
+    """
+    return tf.get_collection(
+        tf.GraphKeys.TRAINABLE_VARIABLES if trainable_only else tf.GraphKeys.GLOBAL_VARIABLES,
+        scope=scope if isinstance(scope, str) else scope.name
+    )
+
+
+def scope_name():
+    """Returns the name of current scope as a string, e.g. deepq/q_func"""
+    return tf.get_variable_scope().name
+
+
+def absolute_scope_name(relative_scope_name):
+    """Appends parent scope name to `relative_scope_name`"""
+    return scope_name() + "/" + relative_scope_name
+
+
+def default_param_noise_filter(var):
+    if var not in tf.trainable_variables():
+        # We never perturb non-trainable vars.
+        return False
+    if "fully_connected" in var.name:
+        # We perturb fully-connected layers.
+        return True
+
+    # The remaining layers are likely conv or layer norm layers, which we do not wish to
+    # perturb (in the former case because they only extract features, in the latter case because
+    # we use them for normalization purposes). If you change your network, you will likely want
+    # to re-consider which layers to perturb and which to keep untouched.
+    return False
+
+
+def build_act(make_obs_ph, q_func, num_actions, scope="deepq", reuse=None):
+    """Creates the act function:
+
+    Parameters
+    ----------
+    make_obs_ph: str -> tf.placeholder or TfInput
+        a function that take a name and creates a placeholder of input with that name
+    q_func: (tf.Variable, int, str, bool) -> tf.Variable
+        the model that takes the following inputs:
+            observation_in: object
+                the output of observation placeholder
+            num_actions: int
+                number of actions
+            scope: str
+            reuse: bool
+                should be passed to outer variable scope
+        and returns a tensor of shape (batch_size, num_actions) with values of every action.
+    num_actions: int
+        number of actions.
+    scope: str or VariableScope
+        optional scope for variable_scope.
+    reuse: bool or None
+        whether or not the variables should be reused. To be able to reuse the scope must be given.
+
+    Returns
+    -------
+    act: (tf.Variable, bool, float) -> tf.Variable
+        function to select and action given observation.
+`       See the top of the file for details.
+    """
+    with tf.variable_scope(scope, reuse=reuse):
+        observations_ph = make_obs_ph("observation")
+        stochastic_ph = tf.placeholder(tf.bool, (), name="stochastic")
+        update_eps_ph = tf.placeholder(tf.float32, (), name="update_eps")
+
+        eps = tf.get_variable("eps", (), initializer=tf.constant_initializer(0))
+
+        q_values = q_func(observations_ph.get(), num_actions, scope="q_func")
+        deterministic_actions = tf.argmax(q_values, axis=1)
+
+        batch_size = tf.shape(observations_ph.get())[0]
+        random_actions = tf.random_uniform(tf.stack([batch_size]), minval=0, maxval=num_actions, dtype=tf.int64)
+        chose_random = tf.random_uniform(tf.stack([batch_size]), minval=0, maxval=1, dtype=tf.float32) < eps
+        stochastic_actions = tf.where(chose_random, random_actions, deterministic_actions)
+
+        output_actions = tf.cond(stochastic_ph, lambda: stochastic_actions, lambda: deterministic_actions)
+        update_eps_expr = eps.assign(tf.cond(update_eps_ph >= 0, lambda: update_eps_ph, lambda: eps))
+        _act = U.function(inputs=[observations_ph, stochastic_ph, update_eps_ph],
+                         outputs=output_actions,
+                         givens={update_eps_ph: -1.0, stochastic_ph: True},
+                         updates=[update_eps_expr])
+        def act(ob, stochastic=True, update_eps=-1):
+            return _act(ob, stochastic, update_eps)
+        return act
+
+
+def build_act_with_param_noise(make_obs_ph, q_func, num_actions, scope="deepq", reuse=None, param_noise_filter_func=None):
+    """Creates the act function with support for parameter space noise exploration (https://arxiv.org/abs/1706.01905):
+
+    Parameters
+    ----------
+    make_obs_ph: str -> tf.placeholder or TfInput
+        a function that take a name and creates a placeholder of input with that name
+    q_func: (tf.Variable, int, str, bool) -> tf.Variable
+        the model that takes the following inputs:
+            observation_in: object
+                the output of observation placeholder
+            num_actions: int
+                number of actions
+            scope: str
+            reuse: bool
+                should be passed to outer variable scope
+        and returns a tensor of shape (batch_size, num_actions) with values of every action.
+    num_actions: int
+        number of actions.
+    scope: str or VariableScope
+        optional scope for variable_scope.
+    reuse: bool or None
+        whether or not the variables should be reused. To be able to reuse the scope must be given.
+    param_noise_filter_func: tf.Variable -> bool
+        function that decides whether or not a variable should be perturbed. Only applicable
+        if param_noise is True. If set to None, default_param_noise_filter is used by default.
+
+    Returns
+    -------
+    act: (tf.Variable, bool, float, bool, float, bool) -> tf.Variable
+        function to select and action given observation.
+`       See the top of the file for details.
+    """
+    if param_noise_filter_func is None:
+        param_noise_filter_func = default_param_noise_filter
+
+    with tf.variable_scope(scope, reuse=reuse):
+        observations_ph = make_obs_ph("observation")
+        stochastic_ph = tf.placeholder(tf.bool, (), name="stochastic")
+        update_eps_ph = tf.placeholder(tf.float32, (), name="update_eps")
+        update_param_noise_threshold_ph = tf.placeholder(tf.float32, (), name="update_param_noise_threshold")
+        update_param_noise_scale_ph = tf.placeholder(tf.bool, (), name="update_param_noise_scale")
+        reset_ph = tf.placeholder(tf.bool, (), name="reset")
+
+        eps = tf.get_variable("eps", (), initializer=tf.constant_initializer(0))
+        param_noise_scale = tf.get_variable("param_noise_scale", (), initializer=tf.constant_initializer(0.01), trainable=False)
+        param_noise_threshold = tf.get_variable("param_noise_threshold", (), initializer=tf.constant_initializer(0.05), trainable=False)
+
+        # Unmodified Q.
+        q_values = q_func(observations_ph.get(), num_actions, scope="q_func")
+
+        # Perturbable Q used for the actual rollout.
+        q_values_perturbed = q_func(observations_ph.get(), num_actions, scope="perturbed_q_func")
+        # We have to wrap this code into a function due to the way tf.cond() works. See
+        # https://stackoverflow.com/questions/37063952/confused-by-the-behavior-of-tf-cond for
+        # a more detailed discussion.
+        def perturb_vars(original_scope, perturbed_scope):
+            all_vars = scope_vars(absolute_scope_name(original_scope))
+            all_perturbed_vars = scope_vars(absolute_scope_name(perturbed_scope))
+            assert len(all_vars) == len(all_perturbed_vars)
+            perturb_ops = []
+            for var, perturbed_var in zip(all_vars, all_perturbed_vars):
+                if param_noise_filter_func(perturbed_var):
+                    # Perturb this variable.
+                    op = tf.assign(perturbed_var, var + tf.random_normal(shape=tf.shape(var), mean=0., stddev=param_noise_scale))
+                else:
+                    # Do not perturb, just assign.
+                    op = tf.assign(perturbed_var, var)
+                perturb_ops.append(op)
+            assert len(perturb_ops) == len(all_vars)
+            return tf.group(*perturb_ops)
+
+        # Set up functionality to re-compute `param_noise_scale`. This perturbs yet another copy
+        # of the network and measures the effect of that perturbation in action space. If the perturbation
+        # is too big, reduce scale of perturbation, otherwise increase.
+        q_values_adaptive = q_func(observations_ph.get(), num_actions, scope="adaptive_q_func")
+        perturb_for_adaption = perturb_vars(original_scope="q_func", perturbed_scope="adaptive_q_func")
+        kl = tf.reduce_sum(tf.nn.softmax(q_values) * (tf.log(tf.nn.softmax(q_values)) - tf.log(tf.nn.softmax(q_values_adaptive))), axis=-1)
+        mean_kl = tf.reduce_mean(kl)
+        def update_scale():
+            with tf.control_dependencies([perturb_for_adaption]):
+                update_scale_expr = tf.cond(mean_kl < param_noise_threshold,
+                    lambda: param_noise_scale.assign(param_noise_scale * 1.01),
+                    lambda: param_noise_scale.assign(param_noise_scale / 1.01),
+                )
+            return update_scale_expr
+
+        # Functionality to update the threshold for parameter space noise.
+        update_param_noise_threshold_expr = param_noise_threshold.assign(tf.cond(update_param_noise_threshold_ph >= 0,
+            lambda: update_param_noise_threshold_ph, lambda: param_noise_threshold))
+
+        # Put everything together.
+        deterministic_actions = tf.argmax(q_values_perturbed, axis=1)
+        batch_size = tf.shape(observations_ph.get())[0]
+        random_actions = tf.random_uniform(tf.stack([batch_size]), minval=0, maxval=num_actions, dtype=tf.int64)
+        chose_random = tf.random_uniform(tf.stack([batch_size]), minval=0, maxval=1, dtype=tf.float32) < eps
+        stochastic_actions = tf.where(chose_random, random_actions, deterministic_actions)
+
+        output_actions = tf.cond(stochastic_ph, lambda: stochastic_actions, lambda: deterministic_actions)
+        update_eps_expr = eps.assign(tf.cond(update_eps_ph >= 0, lambda: update_eps_ph, lambda: eps))
+        updates = [
+            update_eps_expr,
+            tf.cond(reset_ph, lambda: perturb_vars(original_scope="q_func", perturbed_scope="perturbed_q_func"), lambda: tf.group(*[])),
+            tf.cond(update_param_noise_scale_ph, lambda: update_scale(), lambda: tf.Variable(0., trainable=False)),
+            update_param_noise_threshold_expr,
+        ]
+        _act = U.function(inputs=[observations_ph, stochastic_ph, update_eps_ph, reset_ph, update_param_noise_threshold_ph, update_param_noise_scale_ph],
+                         outputs=output_actions,
+                         givens={update_eps_ph: -1.0, stochastic_ph: True, reset_ph: False, update_param_noise_threshold_ph: False, update_param_noise_scale_ph: False},
+                         updates=updates)
+        def act(ob, reset=False, update_param_noise_threshold=False, update_param_noise_scale=False, stochastic=True, update_eps=-1):
+            return _act(ob, stochastic, update_eps, reset, update_param_noise_threshold, update_param_noise_scale)
+        return act
+
+
+def build_train(make_obs_ph, q_func, num_actions, optimizer, grad_norm_clipping=None, gamma=1.0,
+    double_q=True, scope="deepq", reuse=None, param_noise=False, param_noise_filter_func=None):
+    """Creates the train function:
+
+    Parameters
+    ----------
+    make_obs_ph: str -> tf.placeholder or TfInput
+        a function that takes a name and creates a placeholder of input with that name
+    q_func: (tf.Variable, int, str, bool) -> tf.Variable
+        the model that takes the following inputs:
+            observation_in: object
+                the output of observation placeholder
+            num_actions: int
+                number of actions
+            scope: str
+            reuse: bool
+                should be passed to outer variable scope
+        and returns a tensor of shape (batch_size, num_actions) with values of every action.
+    num_actions: int
+        number of actions
+    reuse: bool
+        whether or not to reuse the graph variables
+    optimizer: tf.train.Optimizer
+        optimizer to use for the Q-learning objective.
+    grad_norm_clipping: float or None
+        clip gradient norms to this value. If None no clipping is performed.
+    gamma: float
+        discount rate.
+    double_q: bool
+        if true will use Double Q Learning (https://arxiv.org/abs/1509.06461).
+        In general it is a good idea to keep it enabled.
+    scope: str or VariableScope
+        optional scope for variable_scope.
+    reuse: bool or None
+        whether or not the variables should be reused. To be able to reuse the scope must be given.
+    param_noise: bool
+        whether or not to use parameter space noise (https://arxiv.org/abs/1706.01905)
+    param_noise_filter_func: tf.Variable -> bool
+        function that decides whether or not a variable should be perturbed. Only applicable
+        if param_noise is True. If set to None, default_param_noise_filter is used by default.
+
+    Returns
+    -------
+    act: (tf.Variable, bool, float) -> tf.Variable
+        function to select and action given observation.
+`       See the top of the file for details.
+    train: (object, np.array, np.array, object, np.array, np.array) -> np.array
+        optimize the error in Bellman's equation.
+`       See the top of the file for details.
+    update_target: () -> ()
+        copy the parameters from optimized Q function to the target Q function.
+`       See the top of the file for details.
+    debug: {str: function}
+        a bunch of functions to print debug data like q_values.
+    """
+    if param_noise:
+        act_f = build_act_with_param_noise(make_obs_ph, q_func, num_actions, scope=scope, reuse=reuse,
+            param_noise_filter_func=param_noise_filter_func)
+    else:
+        act_f = build_act(make_obs_ph, q_func, num_actions, scope=scope, reuse=reuse)
+
+    with tf.variable_scope(scope, reuse=reuse):
+        # set up placeholders
+        obs_t_input = make_obs_ph("obs_t")
+        act_t_ph = tf.placeholder(tf.int32, [None], name="action")
+        rew_t_ph = tf.placeholder(tf.float32, [None], name="reward")
+        obs_tp1_input = make_obs_ph("obs_tp1")
+        done_mask_ph = tf.placeholder(tf.float32, [None], name="done")
+        importance_weights_ph = tf.placeholder(tf.float32, [None], name="weight")
+
+        # q network evaluation
+        q_t = q_func(obs_t_input.get(), num_actions, scope="q_func", reuse=True)  # reuse parameters from act
+        q_func_vars = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope=tf.get_variable_scope().name + "/q_func")
+
+        # target q network evalution
+        q_tp1 = q_func(obs_tp1_input.get(), num_actions, scope="target_q_func")
+        target_q_func_vars = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope=tf.get_variable_scope().name + "/target_q_func")
+
+        # q scores for actions which we know were selected in the given state.
+        q_t_selected = tf.reduce_sum(q_t * tf.one_hot(act_t_ph, num_actions), 1)
+
+        # compute estimate of best possible value starting from state at t + 1
+        if double_q:
+            q_tp1_using_online_net = q_func(obs_tp1_input.get(), num_actions, scope="q_func", reuse=True)
+            q_tp1_best_using_online_net = tf.argmax(q_tp1_using_online_net, 1)
+            q_tp1_best = tf.reduce_sum(q_tp1 * tf.one_hot(q_tp1_best_using_online_net, num_actions), 1)
+        else:
+            q_tp1_best = tf.reduce_max(q_tp1, 1)
+        q_tp1_best_masked = (1.0 - done_mask_ph) * q_tp1_best
+
+        # compute RHS of bellman equation
+        q_t_selected_target = rew_t_ph + gamma * q_tp1_best_masked
+
+        # compute the error (potentially clipped)
+        td_error = q_t_selected - tf.stop_gradient(q_t_selected_target)
+        errors = U.huber_loss(td_error)
+        weighted_error = tf.reduce_mean(importance_weights_ph * errors)
+
+        # compute optimization op (potentially with gradient clipping)
+        if grad_norm_clipping is not None:
+            gradients = optimizer.compute_gradients(weighted_error, var_list=q_func_vars)
+            for i, (grad, var) in enumerate(gradients):
+                if grad is not None:
+                    gradients[i] = (tf.clip_by_norm(grad, grad_norm_clipping), var)
+            optimize_expr = optimizer.apply_gradients(gradients)
+        else:
+            optimize_expr = optimizer.minimize(weighted_error, var_list=q_func_vars)
+
+        # update_target_fn will be called periodically to copy Q network to target Q network
+        update_target_expr = []
+        for var, var_target in zip(sorted(q_func_vars, key=lambda v: v.name),
+                                   sorted(target_q_func_vars, key=lambda v: v.name)):
+            update_target_expr.append(var_target.assign(var))
+        update_target_expr = tf.group(*update_target_expr)
+
+        # Create callable functions
+        train = U.function(
+            inputs=[
+                obs_t_input,
+                act_t_ph,
+                rew_t_ph,
+                obs_tp1_input,
+                done_mask_ph,
+                importance_weights_ph
+            ],
+            outputs=td_error,
+            updates=[optimize_expr]
+        )
+        update_target = U.function([], [], updates=[update_target_expr])
+
+        q_values = U.function([obs_t_input], q_t)
+
+        return act_f, train, update_target, {'q_values': q_values}

baselines/deepq/deepq.py

@@ -1,19 +1,96 @@
-import os.path as osp
+import os
+import tempfile
 
 import tensorflow as tf
+import zipfile
+import cloudpickle
 import numpy as np
 
+import baselines.common.tf_util as U
+from baselines.common.tf_util import load_variables, save_variables
 from baselines import logger
 from baselines.common.schedules import LinearSchedule
-from baselines.common.vec_env.vec_env import VecEnv
 from baselines.common import set_global_seeds
 
 from baselines import deepq
 from baselines.deepq.replay_buffer import ReplayBuffer, PrioritizedReplayBuffer
+from baselines.deepq.utils import ObservationInput
 
+from baselines.common.tf_util import get_session
 from baselines.deepq.models import build_q_func
 
 
+class ActWrapper(object):
+    def __init__(self, act, act_params):
+        self._act = act
+        self._act_params = act_params
+        self.initial_state = None
+
+    @staticmethod
+    def load_act(path):
+        with open(path, "rb") as f:
+            model_data, act_params = cloudpickle.load(f)
+        act = deepq.build_act(**act_params)
+        sess = tf.Session()
+        sess.__enter__()
+        with tempfile.TemporaryDirectory() as td:
+            arc_path = os.path.join(td, "packed.zip")
+            with open(arc_path, "wb") as f:
+                f.write(model_data)
+
+            zipfile.ZipFile(arc_path, 'r', zipfile.ZIP_DEFLATED).extractall(td)
+            load_variables(os.path.join(td, "model"))
+
+        return ActWrapper(act, act_params)
+
+    def __call__(self, *args, **kwargs):
+        return self._act(*args, **kwargs)
+
+    def step(self, observation, **kwargs):
+        # DQN doesn't use RNNs so we ignore states and masks
+        kwargs.pop('S', None)
+        kwargs.pop('M', None)
+        return self._act([observation], **kwargs), None, None, None
+
+    def save_act(self, path=None):
+        """Save model to a pickle located at `path`"""
+        if path is None:
+            path = os.path.join(logger.get_dir(), "model.pkl")
+
+        with tempfile.TemporaryDirectory() as td:
+            save_variables(os.path.join(td, "model"))
+            arc_name = os.path.join(td, "packed.zip")
+            with zipfile.ZipFile(arc_name, 'w') as zipf:
+                for root, dirs, files in os.walk(td):
+                    for fname in files:
+                        file_path = os.path.join(root, fname)
+                        if file_path != arc_name:
+                            zipf.write(file_path, os.path.relpath(file_path, td))
+            with open(arc_name, "rb") as f:
+                model_data = f.read()
+        with open(path, "wb") as f:
+            cloudpickle.dump((model_data, self._act_params), f)
+
+    def save(self, path):
+        save_variables(path)
+
+
+def load_act(path):
+    """Load act function that was returned by learn function.
+
+    Parameters
+    ----------
+    path: str
+        path to the act function pickle
+
+    Returns
+    -------
+    act: ActWrapper
+        function that takes a batch of observations
+        and returns actions.
+    """
+    return ActWrapper.load_act(path)
+
 
 def learn(env,
           network,
@@ -110,6 +187,7 @@ def learn(env,
     """
     # Create all the functions necessary to train the model
 
+    sess = get_session()
     set_global_seeds(seed)
 
     q_func = build_q_func(network, **network_kwargs)
@@ -118,23 +196,26 @@ def learn(env,
     # by cloudpickle when serializing make_obs_ph
 
     observation_space = env.observation_space
+    def make_obs_ph(name):
+        return ObservationInput(observation_space, name=name)
 
-    model = deepq.DEEPQ(
+    act, train, update_target, debug = deepq.build_train(
+        make_obs_ph=make_obs_ph,
         q_func=q_func,
-        observation_shape=env.observation_space.shape,
         num_actions=env.action_space.n,
-        lr=lr,
-        grad_norm_clipping=10,
+        optimizer=tf.train.AdamOptimizer(learning_rate=lr),
         gamma=gamma,
+        grad_norm_clipping=10,
         param_noise=param_noise
     )
 
-    if load_path is not None:
-        load_path = osp.expanduser(load_path)
-        ckpt = tf.train.Checkpoint(model=model)
-        manager = tf.train.CheckpointManager(ckpt, load_path, max_to_keep=None)
-        ckpt.restore(manager.latest_checkpoint)
-        print("Restoring from {}".format(manager.latest_checkpoint))
+    act_params = {
+        'make_obs_ph': make_obs_ph,
+        'q_func': q_func,
+        'num_actions': env.action_space.n,
+    }
+
+    act = ActWrapper(act, act_params)
 
     # Create the replay buffer
     if prioritized_replay:
@@ -152,83 +233,101 @@ def learn(env,
                                  initial_p=1.0,
                                  final_p=exploration_final_eps)
 
-    model.update_target()
+    # Initialize the parameters and copy them to the target network.
+    U.initialize()
+    update_target()
 
     episode_rewards = [0.0]
     saved_mean_reward = None
     obs = env.reset()
-    # always mimic the vectorized env
-    if not isinstance(env, VecEnv):
-        obs = np.expand_dims(np.array(obs), axis=0)
     reset = True
 
-    for t in range(total_timesteps):
-        if callback is not None:
-            if callback(locals(), globals()):
-                break
-        kwargs = {}
-        if not param_noise:
-            update_eps = tf.constant(exploration.value(t))
-            update_param_noise_threshold = 0.
-        else:
-            update_eps = tf.constant(0.)
-            # Compute the threshold such that the KL divergence between perturbed and non-perturbed
-            # policy is comparable to eps-greedy exploration with eps = exploration.value(t).
-            # See Appendix C.1 in Parameter Space Noise for Exploration, Plappert et al., 2017
-            # for detailed explanation.
-            update_param_noise_threshold = -np.log(1. - exploration.value(t) + exploration.value(t) / float(env.action_space.n))
-            kwargs['reset'] = reset
-            kwargs['update_param_noise_threshold'] = update_param_noise_threshold
-            kwargs['update_param_noise_scale'] = True
-        action, _, _, _ = model.step(tf.constant(obs), update_eps=update_eps, **kwargs)
-        action = action[0].numpy()
-        reset = False
-        new_obs, rew, done, _ = env.step(action)
-        # Store transition in the replay buffer.
-        if not isinstance(env, VecEnv):
-            new_obs = np.expand_dims(np.array(new_obs), axis=0)
-            replay_buffer.add(obs[0], action, rew, new_obs[0], float(done))
-        else:
-            replay_buffer.add(obs[0], action, rew[0], new_obs[0], float(done[0]))
-        # # Store transition in the replay buffer.
-        # replay_buffer.add(obs, action, rew, new_obs, float(done))
-        obs = new_obs
+    with tempfile.TemporaryDirectory() as td:
+        td = checkpoint_path or td
 
-        episode_rewards[-1] += rew
-        if done:
-            obs = env.reset()
-            if not isinstance(env, VecEnv):
-                obs = np.expand_dims(np.array(obs), axis=0)
-            episode_rewards.append(0.0)
-            reset = True
+        model_file = os.path.join(td, "model")
+        model_saved = False
 
-        if t > learning_starts and t % train_freq == 0:
-            # Minimize the error in Bellman's equation on a batch sampled from replay buffer.
-            if prioritized_replay:
-                experience = replay_buffer.sample(batch_size, beta=beta_schedule.value(t))
-                (obses_t, actions, rewards, obses_tp1, dones, weights, batch_idxes) = experience
+        if tf.train.latest_checkpoint(td) is not None:
+            load_variables(model_file)
+            logger.log('Loaded model from {}'.format(model_file))
+            model_saved = True
+        elif load_path is not None:
+            load_variables(load_path)
+            logger.log('Loaded model from {}'.format(load_path))
+
+
+        for t in range(total_timesteps):
+            if callback is not None:
+                if callback(locals(), globals()):
+                    break
+            # Take action and update exploration to the newest value
+            kwargs = {}
+            if not param_noise:
+                update_eps = exploration.value(t)
+                update_param_noise_threshold = 0.
             else:
-                obses_t, actions, rewards, obses_tp1, dones = replay_buffer.sample(batch_size)
-                weights, batch_idxes = np.ones_like(rewards), None
-            obses_t, obses_tp1 = tf.constant(obses_t), tf.constant(obses_tp1)
-            actions, rewards, dones = tf.constant(actions), tf.constant(rewards), tf.constant(dones)
-            weights = tf.constant(weights)
-            td_errors = model.train(obses_t, actions, rewards, obses_tp1, dones, weights)
-            if prioritized_replay:
-                new_priorities = np.abs(td_errors) + prioritized_replay_eps
-                replay_buffer.update_priorities(batch_idxes, new_priorities)
+                update_eps = 0.
+                # Compute the threshold such that the KL divergence between perturbed and non-perturbed
+                # policy is comparable to eps-greedy exploration with eps = exploration.value(t).
+                # See Appendix C.1 in Parameter Space Noise for Exploration, Plappert et al., 2017
+                # for detailed explanation.
+                update_param_noise_threshold = -np.log(1. - exploration.value(t) + exploration.value(t) / float(env.action_space.n))
+                kwargs['reset'] = reset
+                kwargs['update_param_noise_threshold'] = update_param_noise_threshold
+                kwargs['update_param_noise_scale'] = True
+            action = act(np.array(obs)[None], update_eps=update_eps, **kwargs)[0]
+            env_action = action
+            reset = False
+            new_obs, rew, done, _ = env.step(env_action)
+            # Store transition in the replay buffer.
+            replay_buffer.add(obs, action, rew, new_obs, float(done))
+            obs = new_obs
 
-        if t > learning_starts and t % target_network_update_freq == 0:
-            # Update target network periodically.
-            model.update_target()
+            episode_rewards[-1] += rew
+            if done:
+                obs = env.reset()
+                episode_rewards.append(0.0)
+                reset = True
 
-        mean_100ep_reward = round(np.mean(episode_rewards[-101:-1]), 1)
-        num_episodes = len(episode_rewards)
-        if done and print_freq is not None and len(episode_rewards) % print_freq == 0:
-            logger.record_tabular("steps", t)
-            logger.record_tabular("episodes", num_episodes)
-            logger.record_tabular("mean 100 episode reward", mean_100ep_reward)
-            logger.record_tabular("% time spent exploring", int(100 * exploration.value(t)))
-            logger.dump_tabular()
+            if t > learning_starts and t % train_freq == 0:
+                # Minimize the error in Bellman's equation on a batch sampled from replay buffer.
+                if prioritized_replay:
+                    experience = replay_buffer.sample(batch_size, beta=beta_schedule.value(t))
+                    (obses_t, actions, rewards, obses_tp1, dones, weights, batch_idxes) = experience
+                else:
+                    obses_t, actions, rewards, obses_tp1, dones = replay_buffer.sample(batch_size)
+                    weights, batch_idxes = np.ones_like(rewards), None
+                td_errors = train(obses_t, actions, rewards, obses_tp1, dones, weights)
+                if prioritized_replay:
+                    new_priorities = np.abs(td_errors) + prioritized_replay_eps
+                    replay_buffer.update_priorities(batch_idxes, new_priorities)
 
-    return model
+            if t > learning_starts and t % target_network_update_freq == 0:
+                # Update target network periodically.
+                update_target()
+
+            mean_100ep_reward = round(np.mean(episode_rewards[-101:-1]), 1)
+            num_episodes = len(episode_rewards)
+            if done and print_freq is not None and len(episode_rewards) % print_freq == 0:
+                logger.record_tabular("steps", t)
+                logger.record_tabular("episodes", num_episodes)
+                logger.record_tabular("mean 100 episode reward", mean_100ep_reward)
+                logger.record_tabular("% time spent exploring", int(100 * exploration.value(t)))
+                logger.dump_tabular()
+
+            if (checkpoint_freq is not None and t > learning_starts and
+                    num_episodes > 100 and t % checkpoint_freq == 0):
+                if saved_mean_reward is None or mean_100ep_reward > saved_mean_reward:
+                    if print_freq is not None:
+                        logger.log("Saving model due to mean reward increase: {} -> {}".format(
+                                   saved_mean_reward, mean_100ep_reward))
+                    save_variables(model_file)
+                    model_saved = True
+                    saved_mean_reward = mean_100ep_reward
+        if model_saved:
+            if print_freq is not None:
+                logger.log("Restored model with mean reward: {}".format(saved_mean_reward))
+            load_variables(model_file)
+
+    return act

baselines/deepq/deepq_learner.py

@@ -1,191 +0,0 @@
-"""Deep Q model
-
-The functions in this model:
-
-======= step ========
-
-    Function to chose an action given an observation
-
-    Parameters
-    ----------
-    observation: tensor
-        Observation that can be feed into the output of make_obs_ph
-    stochastic: bool
-        if set to False all the actions are always deterministic (default False)
-    update_eps: float
-        update epsilon a new value, if negative not update happens
-        (default: no update)
-
-    Returns
-    -------
-    Tensor of dtype tf.int64 and shape (BATCH_SIZE,) with an action to be performed for
-    every element of the batch.
-
-
-(NOT IMPLEMENTED YET)
-======= step (in case of parameter noise) ========
-
-    Function to chose an action given an observation
-
-    Parameters
-    ----------
-    observation: object
-        Observation that can be feed into the output of make_obs_ph
-    stochastic: bool
-        if set to False all the actions are always deterministic (default False)
-    update_eps: float
-        update epsilon to a new value, if negative no update happens
-        (default: no update)
-    reset: bool
-        reset the perturbed policy by sampling a new perturbation
-    update_param_noise_threshold: float
-        the desired threshold for the difference between non-perturbed and perturbed policy
-    update_param_noise_scale: bool
-        whether or not to update the scale of the noise for the next time it is re-perturbed
-
-    Returns
-    -------
-    Tensor of dtype tf.int64 and shape (BATCH_SIZE,) with an action to be performed for
-    every element of the batch.
-
-
-======= train =======
-
-    Function that takes a transition (s,a,r,s',d) and optimizes Bellman equation's error:
-
-        td_error = Q(s,a) - (r + gamma * (1-d) * max_a' Q(s', a'))
-        loss = huber_loss[td_error]
-
-    Parameters
-    ----------
-    obs_t: object
-        a batch of observations
-    action: np.array
-        actions that were selected upon seeing obs_t.
-        dtype must be int32 and shape must be (batch_size,)
-    reward: np.array
-        immediate reward attained after executing those actions
-        dtype must be float32 and shape must be (batch_size,)
-    obs_tp1: object
-        observations that followed obs_t
-    done: np.array
-        1 if obs_t was the last observation in the episode and 0 otherwise
-        obs_tp1 gets ignored, but must be of the valid shape.
-        dtype must be float32 and shape must be (batch_size,)
-    weight: np.array
-        imporance weights for every element of the batch (gradient is multiplied
-        by the importance weight) dtype must be float32 and shape must be (batch_size,)
-
-    Returns
-    -------
-    td_error: np.array
-        a list of differences between Q(s,a) and the target in Bellman's equation.
-        dtype is float32 and shape is (batch_size,)
-
-======= update_target ========
-
-    copy the parameters from optimized Q function to the target Q function.
-    In Q learning we actually optimize the following error:
-
-        Q(s,a) - (r + gamma * max_a' Q'(s', a'))
-
-    Where Q' is lagging behind Q to stablize the learning. For example for Atari
-
-    Q' is set to Q once every 10000 updates training steps.
-
-"""
-import tensorflow as tf
-
-@tf.function
-def huber_loss(x, delta=1.0):
-    """Reference: https://en.wikipedia.org/wiki/Huber_loss"""
-    return tf.where(
-        tf.abs(x) < delta,
-        tf.square(x) * 0.5,
-        delta * (tf.abs(x) - 0.5 * delta)
-    )
-
-class DEEPQ(tf.Module):
-
-    def __init__(self, q_func, observation_shape, num_actions, lr, grad_norm_clipping=None, gamma=1.0,
-        double_q=True, param_noise=False, param_noise_filter_func=None):
-
-      self.num_actions = num_actions
-      self.gamma = gamma
-      self.double_q = double_q
-      self.param_noise = param_noise
-      self.param_noise_filter_func = param_noise_filter_func
-      self.grad_norm_clipping = grad_norm_clipping
-
-      self.optimizer = tf.keras.optimizers.Adam(lr)
-
-      with tf.name_scope('q_network'):
-        self.q_network = q_func(observation_shape, num_actions)
-      with tf.name_scope('target_q_network'):
-        self.target_q_network = q_func(observation_shape, num_actions)
-      self.eps = tf.Variable(0., name="eps")
-
-    @tf.function
-    def step(self, obs, stochastic=True, update_eps=-1):
-      if self.param_noise:
-        raise ValueError('not supporting noise yet')
-      else:
-        q_values = self.q_network(obs)
-        deterministic_actions = tf.argmax(q_values, axis=1)
-        batch_size = tf.shape(obs)[0]
-        random_actions = tf.random.uniform(tf.stack([batch_size]), minval=0, maxval=self.num_actions, dtype=tf.int64)
-        chose_random = tf.random.uniform(tf.stack([batch_size]), minval=0, maxval=1, dtype=tf.float32) < self.eps
-        stochastic_actions = tf.where(chose_random, random_actions, deterministic_actions)
-
-        if stochastic:
-          output_actions = stochastic_actions
-        else:
-          output_actions = deterministic_actions
-
-        if update_eps >= 0:
-            self.eps.assign(update_eps)
-
-        return output_actions, None, None, None
-
-    @tf.function()
-    def train(self, obs0, actions, rewards, obs1, dones, importance_weights):
-      with tf.GradientTape() as tape:
-        q_t = self.q_network(obs0)
-        q_t_selected = tf.reduce_sum(q_t * tf.one_hot(actions, self.num_actions, dtype=tf.float32), 1)
-
-        q_tp1 = self.target_q_network(obs1)
-
-        if self.double_q:
-            q_tp1_using_online_net = self.q_network(obs1)
-            q_tp1_best_using_online_net = tf.argmax(q_tp1_using_online_net, 1)
-            q_tp1_best = tf.reduce_sum(q_tp1 * tf.one_hot(q_tp1_best_using_online_net, self.num_actions, dtype=tf.float32), 1)
-        else:
-            q_tp1_best = tf.reduce_max(q_tp1, 1)
-
-        dones = tf.cast(dones, q_tp1_best.dtype)
-        q_tp1_best_masked = (1.0 - dones) * q_tp1_best
-
-        q_t_selected_target = rewards + self.gamma * q_tp1_best_masked
-
-        td_error = q_t_selected - tf.stop_gradient(q_t_selected_target)
-        errors = huber_loss(td_error)
-        weighted_error = tf.reduce_mean(importance_weights * errors)
-
-      grads = tape.gradient(weighted_error, self.q_network.trainable_variables)
-      if self.grad_norm_clipping:
-        clipped_grads = []
-        for grad in grads:
-          clipped_grads.append(tf.clip_by_norm(grad, self.grad_norm_clipping))
-        clipped_grads = grads
-      grads_and_vars = zip(grads, self.q_network.trainable_variables)
-      self.optimizer.apply_gradients(grads_and_vars)
-
-      return td_error
-
-    @tf.function(autograph=False)
-    def update_target(self):
-      q_vars = self.q_network.trainable_variables
-      target_q_vars = self.target_q_network.trainable_variables
-      for var, var_target in zip(q_vars, target_q_vars):
-        var_target.assign(var)
-

baselines/deepq/experiments/custom_cartpole.py

@@ -0,0 +1,79 @@
+import gym
+import itertools
+import numpy as np
+import tensorflow as tf
+import tensorflow.contrib.layers as layers
+
+import baselines.common.tf_util as U
+
+from baselines import logger
+from baselines import deepq
+from baselines.deepq.replay_buffer import ReplayBuffer
+from baselines.deepq.utils import ObservationInput
+from baselines.common.schedules import LinearSchedule
+
+
+def model(inpt, num_actions, scope, reuse=False):
+    """This model takes as input an observation and returns values of all actions."""
+    with tf.variable_scope(scope, reuse=reuse):
+        out = inpt
+        out = layers.fully_connected(out, num_outputs=64, activation_fn=tf.nn.tanh)
+        out = layers.fully_connected(out, num_outputs=num_actions, activation_fn=None)
+        return out
+
+
+if __name__ == '__main__':
+    with U.make_session(num_cpu=8):
+        # Create the environment
+        env = gym.make("CartPole-v0")
+        # Create all the functions necessary to train the model
+        act, train, update_target, debug = deepq.build_train(
+            make_obs_ph=lambda name: ObservationInput(env.observation_space, name=name),
+            q_func=model,
+            num_actions=env.action_space.n,
+            optimizer=tf.train.AdamOptimizer(learning_rate=5e-4),
+        )
+        # Create the replay buffer
+        replay_buffer = ReplayBuffer(50000)
+        # Create the schedule for exploration starting from 1 (every action is random) down to
+        # 0.02 (98% of actions are selected according to values predicted by the model).
+        exploration = LinearSchedule(schedule_timesteps=10000, initial_p=1.0, final_p=0.02)
+
+        # Initialize the parameters and copy them to the target network.
+        U.initialize()
+        update_target()
+
+        episode_rewards = [0.0]
+        obs = env.reset()
+        for t in itertools.count():
+            # Take action and update exploration to the newest value
+            action = act(obs[None], update_eps=exploration.value(t))[0]
+            new_obs, rew, done, _ = env.step(action)
+            # Store transition in the replay buffer.
+            replay_buffer.add(obs, action, rew, new_obs, float(done))
+            obs = new_obs
+
+            episode_rewards[-1] += rew
+            if done:
+                obs = env.reset()
+                episode_rewards.append(0)
+
+            is_solved = t > 100 and np.mean(episode_rewards[-101:-1]) >= 200
+            if is_solved:
+                # Show off the result
+                env.render()
+            else:
+                # Minimize the error in Bellman's equation on a batch sampled from replay buffer.
+                if t > 1000:
+                    obses_t, actions, rewards, obses_tp1, dones = replay_buffer.sample(32)
+                    train(obses_t, actions, rewards, obses_tp1, dones, np.ones_like(rewards))
+                # Update target network periodically.
+                if t % 1000 == 0:
+                    update_target()
+
+            if done and len(episode_rewards) % 10 == 0:
+                logger.record_tabular("steps", t)
+                logger.record_tabular("episodes", len(episode_rewards))
+                logger.record_tabular("mean episode reward", round(np.mean(episode_rewards[-101:-1]), 1))
+                logger.record_tabular("% time spent exploring", int(100 * exploration.value(t)))
+                logger.dump_tabular()

baselines/deepq/experiments/enjoy_cartpole.py

@@ -0,0 +1,21 @@
+import gym
+
+from baselines import deepq
+
+
+def main():
+    env = gym.make("CartPole-v0")
+    act = deepq.learn(env, network='mlp', total_timesteps=0, load_path="cartpole_model.pkl")
+
+    while True:
+        obs, done = env.reset(), False
+        episode_rew = 0
+        while not done:
+            env.render()
+            obs, rew, done, _ = env.step(act(obs[None])[0])
+            episode_rew += rew
+        print("Episode reward", episode_rew)
+
+
+if __name__ == '__main__':
+    main()

baselines/deepq/experiments/enjoy_mountaincar.py

@@ -0,0 +1,27 @@
+import gym
+
+from baselines import deepq
+from baselines.common import models
+
+
+def main():
+    env = gym.make("MountainCar-v0")
+    act = deepq.learn(
+        env,
+        network=models.mlp(num_layers=1, num_hidden=64),
+        total_timesteps=0,
+        load_path='mountaincar_model.pkl'
+    )
+
+    while True:
+        obs, done = env.reset(), False
+        episode_rew = 0
+        while not done:
+            env.render()
+            obs, rew, done, _ = env.step(act(obs[None])[0])
+            episode_rew += rew
+        print("Episode reward", episode_rew)
+
+
+if __name__ == '__main__':
+    main()

baselines/deepq/experiments/enjoy_pong.py

@@ -0,0 +1,28 @@
+import gym
+from baselines import deepq
+
+
+def main():
+    env = gym.make("PongNoFrameskip-v4")
+    env = deepq.wrap_atari_dqn(env)
+    model = deepq.learn(
+        env,
+        "conv_only",
+        convs=[(32, 8, 4), (64, 4, 2), (64, 3, 1)],
+        hiddens=[256],
+        dueling=True,
+        total_timesteps=0
+    )
+
+    while True:
+        obs, done = env.reset(), False
+        episode_rew = 0
+        while not done:
+            env.render()
+            obs, rew, done, _ = env.step(model(obs[None])[0])
+            episode_rew += rew
+        print("Episode reward", episode_rew)
+
+
+if __name__ == '__main__':
+    main()

baselines/deepq/experiments/train_cartpole.py

@@ -0,0 +1,30 @@
+import gym
+
+from baselines import deepq
+
+
+def callback(lcl, _glb):
+    # stop training if reward exceeds 199
+    is_solved = lcl['t'] > 100 and sum(lcl['episode_rewards'][-101:-1]) / 100 >= 199
+    return is_solved
+
+
+def main():
+    env = gym.make("CartPole-v0")
+    act = deepq.learn(
+        env,
+        network='mlp',
+        lr=1e-3,
+        total_timesteps=100000,
+        buffer_size=50000,
+        exploration_fraction=0.1,
+        exploration_final_eps=0.02,
+        print_freq=10,
+        callback=callback
+    )
+    print("Saving model to cartpole_model.pkl")
+    act.save("cartpole_model.pkl")
+
+
+if __name__ == '__main__':
+    main()

baselines/deepq/experiments/train_mountaincar.py

@@ -0,0 +1,26 @@
+import gym
+
+from baselines import deepq
+from baselines.common import models
+
+
+def main():
+    env = gym.make("MountainCar-v0")
+    # Enabling layer_norm here is import for parameter space noise!
+    act = deepq.learn(
+        env,
+        network=models.mlp(num_hidden=64, num_layers=1),
+        lr=1e-3,
+        total_timesteps=100000,
+        buffer_size=50000,
+        exploration_fraction=0.1,
+        exploration_final_eps=0.1,
+        print_freq=10,
+        param_noise=True
+    )
+    print("Saving model to mountaincar_model.pkl")
+    act.save("mountaincar_model.pkl")
+
+
+if __name__ == '__main__':
+    main()

baselines/deepq/experiments/train_pong.py

@@ -0,0 +1,34 @@
+from baselines import deepq
+from baselines import bench
+from baselines import logger
+from baselines.common.atari_wrappers import make_atari
+
+
+def main():
+    logger.configure()
+    env = make_atari('PongNoFrameskip-v4')
+    env = bench.Monitor(env, logger.get_dir())
+    env = deepq.wrap_atari_dqn(env)
+
+    model = deepq.learn(
+        env,
+        "conv_only",
+        convs=[(32, 8, 4), (64, 4, 2), (64, 3, 1)],
+        hiddens=[256],
+        dueling=True,
+        lr=1e-4,
+        total_timesteps=int(1e7),
+        buffer_size=10000,
+        exploration_fraction=0.1,
+        exploration_final_eps=0.01,
+        train_freq=4,
+        learning_starts=10000,
+        target_network_update_freq=1000,
+        gamma=0.99,
+    )
+
+    model.save('pong_model.pkl')
+    env.close()
+
+if __name__ == '__main__':
+    main()

baselines/deepq/models.py

@@ -1,4 +1,100 @@
 import tensorflow as tf
+import tensorflow.contrib.layers as layers
+
+
+def _mlp(hiddens, input_, num_actions, scope, reuse=False, layer_norm=False):
+    with tf.variable_scope(scope, reuse=reuse):
+        out = input_
+        for hidden in hiddens:
+            out = layers.fully_connected(out, num_outputs=hidden, activation_fn=None)
+            if layer_norm:
+                out = layers.layer_norm(out, center=True, scale=True)
+            out = tf.nn.relu(out)
+        q_out = layers.fully_connected(out, num_outputs=num_actions, activation_fn=None)
+        return q_out
+
+
+def mlp(hiddens=[], layer_norm=False):
+    """This model takes as input an observation and returns values of all actions.
+
+    Parameters
+    ----------
+    hiddens: [int]
+        list of sizes of hidden layers
+    layer_norm: bool
+        if true applies layer normalization for every layer
+        as described in https://arxiv.org/abs/1607.06450
+
+    Returns
+    -------
+    q_func: function
+        q_function for DQN algorithm.
+    """
+    return lambda *args, **kwargs: _mlp(hiddens, layer_norm=layer_norm, *args, **kwargs)
+
+
+def _cnn_to_mlp(convs, hiddens, dueling, input_, num_actions, scope, reuse=False, layer_norm=False):
+    with tf.variable_scope(scope, reuse=reuse):
+        out = input_
+        with tf.variable_scope("convnet"):
+            for num_outputs, kernel_size, stride in convs:
+                out = layers.convolution2d(out,
+                                           num_outputs=num_outputs,
+                                           kernel_size=kernel_size,
+                                           stride=stride,
+                                           activation_fn=tf.nn.relu)
+        conv_out = layers.flatten(out)
+        with tf.variable_scope("action_value"):
+            action_out = conv_out
+            for hidden in hiddens:
+                action_out = layers.fully_connected(action_out, num_outputs=hidden, activation_fn=None)
+                if layer_norm:
+                    action_out = layers.layer_norm(action_out, center=True, scale=True)
+                action_out = tf.nn.relu(action_out)
+            action_scores = layers.fully_connected(action_out, num_outputs=num_actions, activation_fn=None)
+
+        if dueling:
+            with tf.variable_scope("state_value"):
+                state_out = conv_out
+                for hidden in hiddens:
+                    state_out = layers.fully_connected(state_out, num_outputs=hidden, activation_fn=None)
+                    if layer_norm:
+                        state_out = layers.layer_norm(state_out, center=True, scale=True)
+                    state_out = tf.nn.relu(state_out)
+                state_score = layers.fully_connected(state_out, num_outputs=1, activation_fn=None)
+            action_scores_mean = tf.reduce_mean(action_scores, 1)
+            action_scores_centered = action_scores - tf.expand_dims(action_scores_mean, 1)
+            q_out = state_score + action_scores_centered
+        else:
+            q_out = action_scores
+        return q_out
+
+
+def cnn_to_mlp(convs, hiddens, dueling=False, layer_norm=False):
+    """This model takes as input an observation and returns values of all actions.
+
+    Parameters
+    ----------
+    convs: [(int, int, int)]
+        list of convolutional layers in form of
+        (num_outputs, kernel_size, stride)
+    hiddens: [int]
+        list of sizes of hidden layers
+    dueling: bool
+        if true double the output MLP to compute a baseline
+        for action scores
+    layer_norm: bool
+        if true applies layer normalization for every layer
+        as described in https://arxiv.org/abs/1607.06450
+
+    Returns
+    -------
+    q_func: function
+        q_function for DQN algorithm.
+    """
+
+    return lambda *args, **kwargs: _cnn_to_mlp(convs, hiddens, dueling, layer_norm=layer_norm, *args, **kwargs)
+
 
 
 def build_q_func(network, hiddens=[256], dueling=True, layer_norm=False, **network_kwargs):
@@ -6,42 +102,39 @@ def build_q_func(network, hiddens=[256], dueling=True, layer_norm=False, **netwo
         from baselines.common.models import get_network_builder
         network = get_network_builder(network)(**network_kwargs)
 
-    def q_func_builder(input_shape, num_actions):
-        # the sub Functional model which does not include the top layer.
-        model = network(input_shape)
+    def q_func_builder(input_placeholder, num_actions, scope, reuse=False):
+        with tf.variable_scope(scope, reuse=reuse):
+            latent = network(input_placeholder)
+            if isinstance(latent, tuple):
+                if latent[1] is not None:
+                    raise NotImplementedError("DQN is not compatible with recurrent policies yet")
+                latent = latent[0]
 
-        # wrapping the sub Functional model with layers that compute action scores into another Functional model.
-        latent = model.outputs
-        if len(latent) > 1:
-            if latent[1] is not None:
-                raise NotImplementedError("DQN is not compatible with recurrent policies yet")
-        latent = latent[0]
+            latent = layers.flatten(latent)
 
-        latent = tf.keras.layers.Flatten()(latent)
-
-        with tf.name_scope("action_value"):
-            action_out = latent
-            for hidden in hiddens:
-                action_out = tf.keras.layers.Dense(units=hidden, activation=None)(action_out)
-                if layer_norm:
-                    action_out = tf.keras.layers.LayerNormalization(center=True, scale=True)(action_out)
-                action_out = tf.nn.relu(action_out)
-            action_scores = tf.keras.layers.Dense(units=num_actions, activation=None)(action_out)
-
-        if dueling:
-            with tf.name_scope("state_value"):
-                state_out = latent
+            with tf.variable_scope("action_value"):
+                action_out = latent
                 for hidden in hiddens:
-                    state_out = tf.keras.layers.Dense(units=hidden, activation=None)(state_out)
+                    action_out = layers.fully_connected(action_out, num_outputs=hidden, activation_fn=None)
                     if layer_norm:
-                        state_out = tf.keras.layers.LayerNormalization(center=True, scale=True)(state_out)
-                    state_out = tf.nn.relu(state_out)
-                state_score = tf.keras.layers.Dense(units=1, activation=None)(state_out)
-            action_scores_mean = tf.reduce_mean(action_scores, 1)
-            action_scores_centered = action_scores - tf.expand_dims(action_scores_mean, 1)
-            q_out = state_score + action_scores_centered
-        else:
-            q_out = action_scores
-        return tf.keras.Model(inputs=model.inputs, outputs=[q_out])
+                        action_out = layers.layer_norm(action_out, center=True, scale=True)
+                    action_out = tf.nn.relu(action_out)
+                action_scores = layers.fully_connected(action_out, num_outputs=num_actions, activation_fn=None)
+
+            if dueling:
+                with tf.variable_scope("state_value"):
+                    state_out = latent
+                    for hidden in hiddens:
+                        state_out = layers.fully_connected(state_out, num_outputs=hidden, activation_fn=None)
+                        if layer_norm:
+                            state_out = layers.layer_norm(state_out, center=True, scale=True)
+                        state_out = tf.nn.relu(state_out)
+                    state_score = layers.fully_connected(state_out, num_outputs=1, activation_fn=None)
+                action_scores_mean = tf.reduce_mean(action_scores, 1)
+                action_scores_centered = action_scores - tf.expand_dims(action_scores_mean, 1)
+                q_out = state_score + action_scores_centered
+            else:
+                q_out = action_scores
+            return q_out
 
     return q_func_builder

baselines/deepq/replay_buffer.py

@@ -32,9 +32,6 @@ class ReplayBuffer(object):
 
     def _encode_sample(self, idxes):
         obses_t, actions, rewards, obses_tp1, dones = [], [], [], [], []
-        data = self._storage[0]
-        ob_dtype = data[0].dtype
-        ac_dtype = data[1].dtype
         for i in idxes:
             data = self._storage[i]
             obs_t, action, reward, obs_tp1, done = data
@@ -43,7 +40,7 @@ class ReplayBuffer(object):
             rewards.append(reward)
             obses_tp1.append(np.array(obs_tp1, copy=False))
             dones.append(done)
-        return np.array(obses_t, dtype=ob_dtype), np.array(actions, dtype=ac_dtype), np.array(rewards, dtype=np.float32), np.array(obses_tp1, dtype=ob_dtype), np.array(dones, dtype=np.float32)
+        return np.array(obses_t), np.array(actions), np.array(rewards), np.array(obses_tp1), np.array(dones)
 
     def sample(self, batch_size):
         """Sample a batch of experiences.
@@ -165,7 +162,7 @@ class PrioritizedReplayBuffer(ReplayBuffer):
             p_sample = self._it_sum[idx] / self._it_sum.sum()
             weight = (p_sample * len(self._storage)) ** (-beta)
             weights.append(weight / max_weight)
-        weights = np.array(weights, dtype=np.float32)
+        weights = np.array(weights)
         encoded_sample = self._encode_sample(idxes)
         return tuple(list(encoded_sample) + [weights, idxes])
 

baselines/deepq/utils.py

@@ -0,0 +1,59 @@
+from baselines.common.input import observation_input
+from baselines.common.tf_util import adjust_shape
+
+# ================================================================
+# Placeholders
+# ================================================================
+
+
+class TfInput(object):
+    def __init__(self, name="(unnamed)"):
+        """Generalized Tensorflow placeholder. The main differences are:
+            - possibly uses multiple placeholders internally and returns multiple values
+            - can apply light postprocessing to the value feed to placeholder.
+        """
+        self.name = name
+
+    def get(self):
+        """Return the tf variable(s) representing the possibly postprocessed value
+        of placeholder(s).
+        """
+        raise NotImplementedError
+
+    def make_feed_dict(self, data):
+        """Given data input it to the placeholder(s)."""
+        raise NotImplementedError
+
+
+class PlaceholderTfInput(TfInput):
+    def __init__(self, placeholder):
+        """Wrapper for regular tensorflow placeholder."""
+        super().__init__(placeholder.name)
+        self._placeholder = placeholder
+
+    def get(self):
+        return self._placeholder
+
+    def make_feed_dict(self, data):
+        return {self._placeholder: adjust_shape(self._placeholder, data)}
+
+
+class ObservationInput(PlaceholderTfInput):
+    def __init__(self, observation_space, name=None):
+        """Creates an input placeholder tailored to a specific observation space
+
+        Parameters
+        ----------
+
+        observation_space:
+                observation space of the environment. Should be one of the gym.spaces types
+        name: str
+                tensorflow name of the underlying placeholder
+        """
+        inpt, self.processed_inpt = observation_input(observation_space, name=name)
+        super().__init__(inpt)
+
+    def get(self):
+        return self.processed_inpt
+
+

baselines/gail/README.md

@@ -0,0 +1,52 @@
+# Generative Adversarial Imitation Learning (GAIL)
+
+- Original paper: https://arxiv.org/abs/1606.03476
+
+For results benchmarking on MuJoCo, please navigate to [here](result/gail-result.md)
+
+## If you want to train an imitation learning agent
+
+### Step 1: Download expert data
+
+Download the expert data into `./data`, [download link](https://drive.google.com/drive/folders/1h3H4AY_ZBx08hz-Ct0Nxxus-V1melu1U?usp=sharing)
+
+### Step 2: Run GAIL
+
+Run with single rank:
+
+```bash
+python -m baselines.gail.run_mujoco
+```
+
+Run with multiple ranks:
+
+```bash
+mpirun -np 16 python -m baselines.gail.run_mujoco
+```
+
+See help (`-h`) for more options.
+
+#### In case you want to run Behavior Cloning (BC)
+
+```bash
+python -m baselines.gail.behavior_clone
+```
+
+See help (`-h`) for more options.
+
+
+## Contributing
+
+Bug reports and pull requests are welcome on GitHub at https://github.com/openai/baselines/pulls.
+
+## Maintainers
+
+- Yuan-Hong Liao, andrewliao11_at_gmail_dot_com
+- Ryan Julian, ryanjulian_at_gmail_dot_com
+
+## Others
+
+Thanks to the open source:
+
+- @openai/imitation
+- @carpedm20/deep-rl-tensorflow

baselines/gail/adversary.py

@@ -0,0 +1,87 @@
+'''
+Reference: https://github.com/openai/imitation
+I follow the architecture from the official repository
+'''
+import tensorflow as tf
+import numpy as np
+
+from baselines.common.mpi_running_mean_std import RunningMeanStd
+from baselines.common import tf_util as U
+
+def logsigmoid(a):
+    '''Equivalent to tf.log(tf.sigmoid(a))'''
+    return -tf.nn.softplus(-a)
+
+""" Reference: https://github.com/openai/imitation/blob/99fbccf3e060b6e6c739bdf209758620fcdefd3c/policyopt/thutil.py#L48-L51"""
+def logit_bernoulli_entropy(logits):
+    ent = (1.-tf.nn.sigmoid(logits))*logits - logsigmoid(logits)
+    return ent
+
+class TransitionClassifier(object):
+    def __init__(self, env, hidden_size, entcoeff=0.001, lr_rate=1e-3, scope="adversary"):
+        self.scope = scope
+        self.observation_shape = env.observation_space.shape
+        self.actions_shape = env.action_space.shape
+        self.input_shape = tuple([o+a for o, a in zip(self.observation_shape, self.actions_shape)])
+        self.num_actions = env.action_space.shape[0]
+        self.hidden_size = hidden_size
+        self.build_ph()
+        # Build grpah
+        generator_logits = self.build_graph(self.generator_obs_ph, self.generator_acs_ph, reuse=False)
+        expert_logits = self.build_graph(self.expert_obs_ph, self.expert_acs_ph, reuse=True)
+        # Build accuracy
+        generator_acc = tf.reduce_mean(tf.to_float(tf.nn.sigmoid(generator_logits) < 0.5))
+        expert_acc = tf.reduce_mean(tf.to_float(tf.nn.sigmoid(expert_logits) > 0.5))
+        # Build regression loss
+        # let x = logits, z = targets.
+        # z * -log(sigmoid(x)) + (1 - z) * -log(1 - sigmoid(x))
+        generator_loss = tf.nn.sigmoid_cross_entropy_with_logits(logits=generator_logits, labels=tf.zeros_like(generator_logits))
+        generator_loss = tf.reduce_mean(generator_loss)
+        expert_loss = tf.nn.sigmoid_cross_entropy_with_logits(logits=expert_logits, labels=tf.ones_like(expert_logits))
+        expert_loss = tf.reduce_mean(expert_loss)
+        # Build entropy loss
+        logits = tf.concat([generator_logits, expert_logits], 0)
+        entropy = tf.reduce_mean(logit_bernoulli_entropy(logits))
+        entropy_loss = -entcoeff*entropy
+        # Loss + Accuracy terms
+        self.losses = [generator_loss, expert_loss, entropy, entropy_loss, generator_acc, expert_acc]
+        self.loss_name = ["generator_loss", "expert_loss", "entropy", "entropy_loss", "generator_acc", "expert_acc"]
+        self.total_loss = generator_loss + expert_loss + entropy_loss
+        # Build Reward for policy
+        self.reward_op = -tf.log(1-tf.nn.sigmoid(generator_logits)+1e-8)
+        var_list = self.get_trainable_variables()
+        self.lossandgrad = U.function([self.generator_obs_ph, self.generator_acs_ph, self.expert_obs_ph, self.expert_acs_ph],
+                                      self.losses + [U.flatgrad(self.total_loss, var_list)])
+
+    def build_ph(self):
+        self.generator_obs_ph = tf.placeholder(tf.float32, (None, ) + self.observation_shape, name="observations_ph")
+        self.generator_acs_ph = tf.placeholder(tf.float32, (None, ) + self.actions_shape, name="actions_ph")
+        self.expert_obs_ph = tf.placeholder(tf.float32, (None, ) + self.observation_shape, name="expert_observations_ph")
+        self.expert_acs_ph = tf.placeholder(tf.float32, (None, ) + self.actions_shape, name="expert_actions_ph")
+
+    def build_graph(self, obs_ph, acs_ph, reuse=False):
+        with tf.variable_scope(self.scope):
+            if reuse:
+                tf.get_variable_scope().reuse_variables()
+
+            with tf.variable_scope("obfilter"):
+                self.obs_rms = RunningMeanStd(shape=self.observation_shape)
+            obs = (obs_ph - self.obs_rms.mean) / self.obs_rms.std
+            _input = tf.concat([obs, acs_ph], axis=1)  # concatenate the two input -> form a transition
+            p_h1 = tf.contrib.layers.fully_connected(_input, self.hidden_size, activation_fn=tf.nn.tanh)
+            p_h2 = tf.contrib.layers.fully_connected(p_h1, self.hidden_size, activation_fn=tf.nn.tanh)
+            logits = tf.contrib.layers.fully_connected(p_h2, 1, activation_fn=tf.identity)
+        return logits
+
+    def get_trainable_variables(self):
+        return tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, self.scope)
+
+    def get_reward(self, obs, acs):
+        sess = tf.get_default_session()
+        if len(obs.shape) == 1:
+            obs = np.expand_dims(obs, 0)
+        if len(acs.shape) == 1:
+            acs = np.expand_dims(acs, 0)
+        feed_dict = {self.generator_obs_ph: obs, self.generator_acs_ph: acs}
+        reward = sess.run(self.reward_op, feed_dict)
+        return reward

baselines/gail/behavior_clone.py

@@ -0,0 +1,124 @@
+'''
+The code is used to train BC imitator, or pretrained GAIL imitator
+'''
+
+import argparse
+import tempfile
+import os.path as osp
+import gym
+import logging
+from tqdm import tqdm
+
+import tensorflow as tf
+
+from baselines.gail import mlp_policy
+from baselines import bench
+from baselines import logger
+from baselines.common import set_global_seeds, tf_util as U
+from baselines.common.misc_util import boolean_flag
+from baselines.common.mpi_adam import MpiAdam
+from baselines.gail.run_mujoco import runner
+from baselines.gail.dataset.mujoco_dset import Mujoco_Dset
+
+
+def argsparser():
+    parser = argparse.ArgumentParser("Tensorflow Implementation of Behavior Cloning")
+    parser.add_argument('--env_id', help='environment ID', default='Hopper-v1')
+    parser.add_argument('--seed', help='RNG seed', type=int, default=0)
+    parser.add_argument('--expert_path', type=str, default='data/deterministic.trpo.Hopper.0.00.npz')
+    parser.add_argument('--checkpoint_dir', help='the directory to save model', default='checkpoint')
+    parser.add_argument('--log_dir', help='the directory to save log file', default='log')
+    #  Mujoco Dataset Configuration
+    parser.add_argument('--traj_limitation', type=int, default=-1)
+    # Network Configuration (Using MLP Policy)
+    parser.add_argument('--policy_hidden_size', type=int, default=100)
+    # for evaluatation
+    boolean_flag(parser, 'stochastic_policy', default=False, help='use stochastic/deterministic policy to evaluate')
+    boolean_flag(parser, 'save_sample', default=False, help='save the trajectories or not')
+    parser.add_argument('--BC_max_iter', help='Max iteration for training BC', type=int, default=1e5)
+    return parser.parse_args()
+
+
+def learn(env, policy_func, dataset, optim_batch_size=128, max_iters=1e4,
+          adam_epsilon=1e-5, optim_stepsize=3e-4,
+          ckpt_dir=None, log_dir=None, task_name=None,
+          verbose=False):
+
+    val_per_iter = int(max_iters/10)
+    ob_space = env.observation_space
+    ac_space = env.action_space
+    pi = policy_func("pi", ob_space, ac_space)  # Construct network for new policy
+    # placeholder
+    ob = U.get_placeholder_cached(name="ob")
+    ac = pi.pdtype.sample_placeholder([None])
+    stochastic = U.get_placeholder_cached(name="stochastic")
+    loss = tf.reduce_mean(tf.square(ac-pi.ac))
+    var_list = pi.get_trainable_variables()
+    adam = MpiAdam(var_list, epsilon=adam_epsilon)
+    lossandgrad = U.function([ob, ac, stochastic], [loss]+[U.flatgrad(loss, var_list)])
+
+    U.initialize()
+    adam.sync()
+    logger.log("Pretraining with Behavior Cloning...")
+    for iter_so_far in tqdm(range(int(max_iters))):
+        ob_expert, ac_expert = dataset.get_next_batch(optim_batch_size, 'train')
+        train_loss, g = lossandgrad(ob_expert, ac_expert, True)
+        adam.update(g, optim_stepsize)
+        if verbose and iter_so_far % val_per_iter == 0:
+            ob_expert, ac_expert = dataset.get_next_batch(-1, 'val')
+            val_loss, _ = lossandgrad(ob_expert, ac_expert, True)
+            logger.log("Training loss: {}, Validation loss: {}".format(train_loss, val_loss))
+
+    if ckpt_dir is None:
+        savedir_fname = tempfile.TemporaryDirectory().name
+    else:
+        savedir_fname = osp.join(ckpt_dir, task_name)
+    U.save_state(savedir_fname, var_list=pi.get_variables())
+    return savedir_fname
+
+
+def get_task_name(args):
+    task_name = 'BC'
+    task_name += '.{}'.format(args.env_id.split("-")[0])
+    task_name += '.traj_limitation_{}'.format(args.traj_limitation)
+    task_name += ".seed_{}".format(args.seed)
+    return task_name
+
+
+def main(args):
+    U.make_session(num_cpu=1).__enter__()
+    set_global_seeds(args.seed)
+    env = gym.make(args.env_id)
+
+    def policy_fn(name, ob_space, ac_space, reuse=False):
+        return mlp_policy.MlpPolicy(name=name, ob_space=ob_space, ac_space=ac_space,
+                                    reuse=reuse, hid_size=args.policy_hidden_size, num_hid_layers=2)
+    env = bench.Monitor(env, logger.get_dir() and
+                        osp.join(logger.get_dir(), "monitor.json"))
+    env.seed(args.seed)
+    gym.logger.setLevel(logging.WARN)
+    task_name = get_task_name(args)
+    args.checkpoint_dir = osp.join(args.checkpoint_dir, task_name)
+    args.log_dir = osp.join(args.log_dir, task_name)
+    dataset = Mujoco_Dset(expert_path=args.expert_path, traj_limitation=args.traj_limitation)
+    savedir_fname = learn(env,
+                          policy_fn,
+                          dataset,
+                          max_iters=args.BC_max_iter,
+                          ckpt_dir=args.checkpoint_dir,
+                          log_dir=args.log_dir,
+                          task_name=task_name,
+                          verbose=True)
+    avg_len, avg_ret = runner(env,
+                              policy_fn,
+                              savedir_fname,
+                              timesteps_per_batch=1024,
+                              number_trajs=10,
+                              stochastic_policy=args.stochastic_policy,
+                              save=args.save_sample,
+                              reuse=True)
+
+
+if __name__ == '__main__':
+    args = argsparser()
+    main(args)

baselines/gail/dataset/mujoco_dset.py

@@ -0,0 +1,114 @@
+'''
+Data structure of the input .npz:
+the data is save in python dictionary format with keys: 'acs', 'ep_rets', 'rews', 'obs'
+the values of each item is a list storing the expert trajectory sequentially
+a transition can be: (data['obs'][t], data['acs'][t], data['obs'][t+1]) and get reward data['rews'][t]
+'''
+
+from baselines import logger
+import numpy as np
+
+
+class Dset(object):
+    def __init__(self, inputs, labels, randomize):
+        self.inputs = inputs
+        self.labels = labels
+        assert len(self.inputs) == len(self.labels)
+        self.randomize = randomize
+        self.num_pairs = len(inputs)
+        self.init_pointer()
+
+    def init_pointer(self):
+        self.pointer = 0
+        if self.randomize:
+            idx = np.arange(self.num_pairs)
+            np.random.shuffle(idx)
+            self.inputs = self.inputs[idx, :]
+            self.labels = self.labels[idx, :]
+
+    def get_next_batch(self, batch_size):
+        # if batch_size is negative -> return all
+        if batch_size < 0:
+            return self.inputs, self.labels
+        if self.pointer + batch_size >= self.num_pairs:
+            self.init_pointer()
+        end = self.pointer + batch_size
+        inputs = self.inputs[self.pointer:end, :]
+        labels = self.labels[self.pointer:end, :]
+        self.pointer = end
+        return inputs, labels
+
+
+class Mujoco_Dset(object):
+    def __init__(self, expert_path, train_fraction=0.7, traj_limitation=-1, randomize=True):
+        traj_data = np.load(expert_path)
+        if traj_limitation < 0:
+            traj_limitation = len(traj_data['obs'])
+        obs = traj_data['obs'][:traj_limitation]
+        acs = traj_data['acs'][:traj_limitation]
+
+        # obs, acs: shape (N, L, ) + S where N = # episodes, L = episode length
+        # and S is the environment observation/action space.
+        # Flatten to (N * L, prod(S))
+        if len(obs.shape) > 2:
+            self.obs = np.reshape(obs, [-1, np.prod(obs.shape[2:])])
+            self.acs = np.reshape(acs, [-1, np.prod(acs.shape[2:])])
+        else:
+            self.obs = np.vstack(obs)
+            self.acs = np.vstack(acs)
+
+        self.rets = traj_data['ep_rets'][:traj_limitation]
+        self.avg_ret = sum(self.rets)/len(self.rets)
+        self.std_ret = np.std(np.array(self.rets))
+        if len(self.acs) > 2:
+            self.acs = np.squeeze(self.acs)
+        assert len(self.obs) == len(self.acs)
+        self.num_traj = min(traj_limitation, len(traj_data['obs']))
+        self.num_transition = len(self.obs)
+        self.randomize = randomize
+        self.dset = Dset(self.obs, self.acs, self.randomize)
+        # for behavior cloning
+        self.train_set = Dset(self.obs[:int(self.num_transition*train_fraction), :],
+                              self.acs[:int(self.num_transition*train_fraction), :],
+                              self.randomize)
+        self.val_set = Dset(self.obs[int(self.num_transition*train_fraction):, :],
+                            self.acs[int(self.num_transition*train_fraction):, :],
+                            self.randomize)
+        self.log_info()
+
+    def log_info(self):
+        logger.log("Total trajectorues: %d" % self.num_traj)
+        logger.log("Total transitions: %d" % self.num_transition)
+        logger.log("Average returns: %f" % self.avg_ret)
+        logger.log("Std for returns: %f" % self.std_ret)
+
+    def get_next_batch(self, batch_size, split=None):
+        if split is None:
+            return self.dset.get_next_batch(batch_size)
+        elif split == 'train':
+            return self.train_set.get_next_batch(batch_size)
+        elif split == 'val':
+            return self.val_set.get_next_batch(batch_size)
+        else:
+            raise NotImplementedError
+
+    def plot(self):
+        import matplotlib.pyplot as plt
+        plt.hist(self.rets)
+        plt.savefig("histogram_rets.png")
+        plt.close()
+
+
+def test(expert_path, traj_limitation, plot):
+    dset = Mujoco_Dset(expert_path, traj_limitation=traj_limitation)
+    if plot:
+        dset.plot()
+
+if __name__ == '__main__':
+    import argparse
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--expert_path", type=str, default="../data/deterministic.trpo.Hopper.0.00.npz")
+    parser.add_argument("--traj_limitation", type=int, default=None)
+    parser.add_argument("--plot", type=bool, default=False)
+    args = parser.parse_args()
+    test(args.expert_path, args.traj_limitation, args.plot)

baselines/gail/gail-eval.py

@@ -0,0 +1,147 @@
+'''
+This code is used to evalaute the imitators trained with different number of trajectories
+and plot the results in the same figure for easy comparison.
+'''
+
+import argparse
+import os
+import glob
+import gym
+
+import matplotlib.pyplot as plt
+import numpy as np
+import tensorflow as tf
+
+from baselines.gail import run_mujoco
+from baselines.gail import mlp_policy
+from baselines.common import set_global_seeds, tf_util as U
+from baselines.common.misc_util import boolean_flag
+from baselines.gail.dataset.mujoco_dset import Mujoco_Dset
+
+
+plt.style.use('ggplot')
+CONFIG = {
+    'traj_limitation': [1, 5, 10, 50],
+}
+
+
+def load_dataset(expert_path):
+    dataset = Mujoco_Dset(expert_path=expert_path)
+    return dataset
+
+
+def argsparser():
+    parser = argparse.ArgumentParser('Do evaluation')
+    parser.add_argument('--seed', type=int, default=0)
+    parser.add_argument('--policy_hidden_size', type=int, default=100)
+    parser.add_argument('--env', type=str, choices=['Hopper', 'Walker2d', 'HalfCheetah',
+                                                    'Humanoid', 'HumanoidStandup'])
+    boolean_flag(parser, 'stochastic_policy', default=False, help='use stochastic/deterministic policy to evaluate')
+    return parser.parse_args()
+
+
+def evaluate_env(env_name, seed, policy_hidden_size, stochastic, reuse, prefix):
+
+    def get_checkpoint_dir(checkpoint_list, limit, prefix):
+        for checkpoint in checkpoint_list:
+            if ('limitation_'+str(limit) in checkpoint) and (prefix in checkpoint):
+                return checkpoint
+        return None
+
+    def policy_fn(name, ob_space, ac_space, reuse=False):
+        return mlp_policy.MlpPolicy(name=name, ob_space=ob_space, ac_space=ac_space,
+                                    reuse=reuse, hid_size=policy_hidden_size, num_hid_layers=2)
+
+    data_path = os.path.join('data', 'deterministic.trpo.' + env_name + '.0.00.npz')
+    dataset = load_dataset(data_path)
+    checkpoint_list = glob.glob(os.path.join('checkpoint', '*' + env_name + ".*"))
+    log = {
+        'traj_limitation': [],
+        'upper_bound': [],
+        'avg_ret': [],
+        'avg_len': [],
+        'normalized_ret': []
+    }
+    for i, limit in enumerate(CONFIG['traj_limitation']):
+        # Do one evaluation
+        upper_bound = sum(dataset.rets[:limit])/limit
+        checkpoint_dir = get_checkpoint_dir(checkpoint_list, limit, prefix=prefix)
+        checkpoint_path = tf.train.latest_checkpoint(checkpoint_dir)
+        env = gym.make(env_name + '-v1')
+        env.seed(seed)
+        print('Trajectory limitation: {}, Load checkpoint: {}, '.format(limit, checkpoint_path))
+        avg_len, avg_ret = run_mujoco.runner(env,
+                                             policy_fn,
+                                             checkpoint_path,
+                                             timesteps_per_batch=1024,
+                                             number_trajs=10,
+                                             stochastic_policy=stochastic,
+                                             reuse=((i != 0) or reuse))
+        normalized_ret = avg_ret/upper_bound
+        print('Upper bound: {}, evaluation returns: {}, normalized scores: {}'.format(
+            upper_bound, avg_ret, normalized_ret))
+        log['traj_limitation'].append(limit)
+        log['upper_bound'].append(upper_bound)
+        log['avg_ret'].append(avg_ret)
+        log['avg_len'].append(avg_len)
+        log['normalized_ret'].append(normalized_ret)
+        env.close()
+    return log
+
+
+def plot(env_name, bc_log, gail_log, stochastic):
+    upper_bound = bc_log['upper_bound']
+    bc_avg_ret = bc_log['avg_ret']
+    gail_avg_ret = gail_log['avg_ret']
+    plt.plot(CONFIG['traj_limitation'], upper_bound)
+    plt.plot(CONFIG['traj_limitation'], bc_avg_ret)
+    plt.plot(CONFIG['traj_limitation'], gail_avg_ret)
+    plt.xlabel('Number of expert trajectories')
+    plt.ylabel('Accumulated reward')
+    plt.title('{} unnormalized scores'.format(env_name))
+    plt.legend(['expert', 'bc-imitator', 'gail-imitator'], loc='lower right')
+    plt.grid(b=True, which='major', color='gray', linestyle='--')
+    if stochastic:
+        title_name = 'result/{}-unnormalized-stochastic-scores.png'.format(env_name)
+    else:
+        title_name = 'result/{}-unnormalized-deterministic-scores.png'.format(env_name)
+    plt.savefig(title_name)
+    plt.close()
+
+    bc_normalized_ret = bc_log['normalized_ret']
+    gail_normalized_ret = gail_log['normalized_ret']
+    plt.plot(CONFIG['traj_limitation'], np.ones(len(CONFIG['traj_limitation'])))
+    plt.plot(CONFIG['traj_limitation'], bc_normalized_ret)
+    plt.plot(CONFIG['traj_limitation'], gail_normalized_ret)
+    plt.xlabel('Number of expert trajectories')
+    plt.ylabel('Normalized performance')
+    plt.title('{} normalized scores'.format(env_name))
+    plt.legend(['expert', 'bc-imitator', 'gail-imitator'], loc='lower right')
+    plt.grid(b=True, which='major', color='gray', linestyle='--')
+    if stochastic:
+        title_name = 'result/{}-normalized-stochastic-scores.png'.format(env_name)
+    else:
+        title_name = 'result/{}-normalized-deterministic-scores.png'.format(env_name)
+    plt.ylim(0, 1.6)
+    plt.savefig(title_name)
+    plt.close()
+
+
+def main(args):
+    U.make_session(num_cpu=1).__enter__()
+    set_global_seeds(args.seed)
+    print('Evaluating {}'.format(args.env))
+    bc_log = evaluate_env(args.env, args.seed, args.policy_hidden_size,
+                          args.stochastic_policy, False, 'BC')
+    print('Evaluation for {}'.format(args.env))
+    print(bc_log)
+    gail_log = evaluate_env(args.env, args.seed, args.policy_hidden_size,
+                            args.stochastic_policy, True, 'gail')
+    print('Evaluation for {}'.format(args.env))
+    print(gail_log)
+    plot(args.env, bc_log, gail_log, args.stochastic_policy)
+
+
+if __name__ == '__main__':
+    args = argsparser()
+    main(args)

baselines/gail/mlp_policy.py

@@ -0,0 +1,75 @@
+'''
+from baselines/ppo1/mlp_policy.py and add simple modification
+(1) add reuse argument
+(2) cache the `stochastic` placeholder
+'''
+import tensorflow as tf
+import gym
+
+import baselines.common.tf_util as U
+from baselines.common.mpi_running_mean_std import RunningMeanStd
+from baselines.common.distributions import make_pdtype
+from baselines.acktr.utils import dense
+
+
+class MlpPolicy(object):
+    recurrent = False
+
+    def __init__(self, name, reuse=False, *args, **kwargs):
+        with tf.variable_scope(name):
+            if reuse:
+                tf.get_variable_scope().reuse_variables()
+            self._init(*args, **kwargs)
+            self.scope = tf.get_variable_scope().name
+
+    def _init(self, ob_space, ac_space, hid_size, num_hid_layers, gaussian_fixed_var=True):
+        assert isinstance(ob_space, gym.spaces.Box)
+
+        self.pdtype = pdtype = make_pdtype(ac_space)
+        sequence_length = None
+
+        ob = U.get_placeholder(name="ob", dtype=tf.float32, shape=[sequence_length] + list(ob_space.shape))
+
+        with tf.variable_scope("obfilter"):
+            self.ob_rms = RunningMeanStd(shape=ob_space.shape)
+
+        obz = tf.clip_by_value((ob - self.ob_rms.mean) / self.ob_rms.std, -5.0, 5.0)
+        last_out = obz
+        for i in range(num_hid_layers):
+            last_out = tf.nn.tanh(dense(last_out, hid_size, "vffc%i" % (i+1), weight_init=U.normc_initializer(1.0)))
+        self.vpred = dense(last_out, 1, "vffinal", weight_init=U.normc_initializer(1.0))[:, 0]
+
+        last_out = obz
+        for i in range(num_hid_layers):
+            last_out = tf.nn.tanh(dense(last_out, hid_size, "polfc%i" % (i+1), weight_init=U.normc_initializer(1.0)))
+
+        if gaussian_fixed_var and isinstance(ac_space, gym.spaces.Box):
+            mean = dense(last_out, pdtype.param_shape()[0]//2, "polfinal", U.normc_initializer(0.01))
+            logstd = tf.get_variable(name="logstd", shape=[1, pdtype.param_shape()[0]//2], initializer=tf.zeros_initializer())
+            pdparam = tf.concat([mean, mean * 0.0 + logstd], axis=1)
+        else:
+            pdparam = dense(last_out, pdtype.param_shape()[0], "polfinal", U.normc_initializer(0.01))
+
+        self.pd = pdtype.pdfromflat(pdparam)
+
+        self.state_in = []
+        self.state_out = []
+
+        # change for BC
+        stochastic = U.get_placeholder(name="stochastic", dtype=tf.bool, shape=())
+        ac = U.switch(stochastic, self.pd.sample(), self.pd.mode())
+        self.ac = ac
+        self._act = U.function([stochastic, ob], [ac, self.vpred])
+
+    def act(self, stochastic, ob):
+        ac1, vpred1 = self._act(stochastic, ob[None])
+        return ac1[0], vpred1[0]
+
+    def get_variables(self):
+        return tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, self.scope)
+
+    def get_trainable_variables(self):
+        return tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, self.scope)
+
+    def get_initial_state(self):
+        return []