remove ref to simple_bench + rank variable not used [mpi gone]

some logger changes

.gitignore

@@ -30,3 +30,6 @@ src
 
 *.egg-info
 .cache
+
+MUJOCO_LOG.TXT
+

README.md

@@ -9,9 +9,25 @@ These algorithms will make it easier for the research community to replicate, re
 You can install it by typing:
 
 ```bash
-pip install baselines
+git clone https://github.com/openai/baselines.git
+cd baselines
+pip install -e .
 ```
 
+- [A2C](baselines/a2c)
+- [ACKTR](baselines/acktr)
+- [DDPG](baselines/ddpg)
 - [DQN](baselines/deepq)
-- [PPO](baselines/pposgd)
+- [PPO](baselines/ppo1)
 - [TRPO](baselines/trpo_mpi)
+
+To cite this repository in publications:
+
+    @misc{baselines,
+      author = {Hesse, Christopher and Plappert, Matthias and Radford, Alec and Schulman, John and Sidor, Szymon and Wu, Yuhuai},
+      title = {OpenAI Baselines},
+      year = {2017},
+      publisher = {GitHub},
+      journal = {GitHub repository},
+      howpublished = {\url{https://github.com/openai/baselines}},
+    }

baselines/a2c/README.md

@@ -0,0 +1,5 @@
+# A2C
+
+- Original paper: https://arxiv.org/abs/1602.01783
+- Baselines blog post: https://blog.openai.com/baselines-acktr-a2c/
+- `python -m baselines.a2c.run_atari` runs the algorithm for 40M frames = 10M timesteps on an Atari game. See help (`-h`) for more options.

baselines/a2c/a2c.py

@@ -0,0 +1,188 @@
+import os.path as osp
+import gym
+import time
+import joblib
+import logging
+import numpy as np
+import tensorflow as tf
+from baselines import logger
+
+from baselines.common import set_global_seeds, explained_variance
+from baselines.common.vec_env.subproc_vec_env import SubprocVecEnv
+from baselines.common.atari_wrappers import wrap_deepmind
+
+from baselines.a2c.utils import discount_with_dones
+from baselines.a2c.utils import Scheduler, make_path, find_trainable_variables
+from baselines.a2c.policies import CnnPolicy
+from baselines.a2c.utils import cat_entropy, mse
+
+class Model(object):
+
+    def __init__(self, policy, ob_space, ac_space, nenvs, nsteps, nstack, num_procs,
+            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), lrschedule='linear'):
+        config = tf.ConfigProto(allow_soft_placement=True,
+                                intra_op_parallelism_threads=num_procs,
+                                inter_op_parallelism_threads=num_procs)
+        config.gpu_options.allow_growth = True
+        sess = tf.Session(config=config)
+        nact = ac_space.n
+        nbatch = nenvs*nsteps
+
+        A = tf.placeholder(tf.int32, [nbatch])
+        ADV = tf.placeholder(tf.float32, [nbatch])
+        R = tf.placeholder(tf.float32, [nbatch])
+        LR = tf.placeholder(tf.float32, [])
+
+        step_model = policy(sess, ob_space, ac_space, nenvs, 1, nstack, reuse=False)
+        train_model = policy(sess, ob_space, ac_space, nenvs, nsteps, nstack, reuse=True)
+
+        neglogpac = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=train_model.pi, labels=A)
+        pg_loss = tf.reduce_mean(ADV * neglogpac)
+        vf_loss = tf.reduce_mean(mse(tf.squeeze(train_model.vf), R))
+        entropy = tf.reduce_mean(cat_entropy(train_model.pi))
+        loss = pg_loss - entropy*ent_coef + vf_loss * vf_coef
+
+        params = find_trainable_variables("model")
+        grads = tf.gradients(loss, params)
+        if max_grad_norm is not None:
+            grads, grad_norm = tf.clip_by_global_norm(grads, max_grad_norm)
+        grads = list(zip(grads, params))
+        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):
+            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 != []:
+                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
+
+        def save(save_path):
+            ps = sess.run(params)
+            make_path(save_path)
+            joblib.dump(ps, save_path)
+
+        def load(load_path):
+            loaded_params = joblib.load(load_path)
+            restores = []
+            for p, loaded_p in zip(params, loaded_params):
+                restores.append(p.assign(loaded_p))
+            ps = sess.run(restores)
+
+        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 = save
+        self.load = load
+        tf.global_variables_initializer().run(session=sess)
+
+class Runner(object):
+
+    def __init__(self, env, model, nsteps=5, nstack=4, gamma=0.99):
+        self.env = env
+        self.model = model
+        nh, nw, nc = env.observation_space.shape
+        nenv = env.num_envs
+        self.batch_ob_shape = (nenv*nsteps, nh, nw, nc*nstack)
+        self.obs = np.zeros((nenv, nh, nw, nc*nstack), dtype=np.uint8)
+        self.nc = nc
+        obs = env.reset()
+        self.update_obs(obs)
+        self.gamma = gamma
+        self.nsteps = nsteps
+        self.states = model.initial_state
+        self.dones = [False for _ in range(nenv)]
+
+    def update_obs(self, obs):
+        # Do frame-stacking here instead of the FrameStack wrapper to reduce
+        # IPC overhead
+        self.obs = np.roll(self.obs, shift=-self.nc, axis=3)
+        self.obs[:, :, :, -self.nc:] = obs
+
+    def run(self):
+        mb_obs, mb_rewards, mb_actions, mb_values, mb_dones = [],[],[],[],[]
+        mb_states = self.states
+        for n in range(self.nsteps):
+            actions, values, states = self.model.step(self.obs, self.states, self.dones)
+            mb_obs.append(np.copy(self.obs))
+            mb_actions.append(actions)
+            mb_values.append(values)
+            mb_dones.append(self.dones)
+            obs, rewards, dones, _ = self.env.step(actions)
+            self.states = states
+            self.dones = dones
+            for n, done in enumerate(dones):
+                if done:
+                    self.obs[n] = self.obs[n]*0
+            self.update_obs(obs)
+            mb_rewards.append(rewards)
+        mb_dones.append(self.dones)
+        #batch of steps to batch of rollouts
+        mb_obs = np.asarray(mb_obs, dtype=np.uint8).swapaxes(1, 0).reshape(self.batch_ob_shape)
+        mb_rewards = np.asarray(mb_rewards, dtype=np.float32).swapaxes(1, 0)
+        mb_actions = np.asarray(mb_actions, dtype=np.int32).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]
+        mb_dones = mb_dones[:, 1:]
+        last_values = self.model.value(self.obs, self.states, self.dones).tolist()
+        #discount/bootstrap off value fn
+        for n, (rewards, dones, value) in enumerate(zip(mb_rewards, mb_dones, last_values)):
+            rewards = rewards.tolist()
+            dones = dones.tolist()
+            if dones[-1] == 0:
+                rewards = discount_with_dones(rewards+[value], dones+[0], self.gamma)[:-1]
+            else:
+                rewards = discount_with_dones(rewards, dones, self.gamma)
+            mb_rewards[n] = rewards
+        mb_rewards = mb_rewards.flatten()
+        mb_actions = mb_actions.flatten()
+        mb_values = mb_values.flatten()
+        mb_masks = mb_masks.flatten()
+        return mb_obs, mb_states, mb_rewards, mb_masks, mb_actions, mb_values
+
+def learn(policy, env, seed, nsteps=5, nstack=4, total_timesteps=int(80e6), vf_coef=0.5, ent_coef=0.01, max_grad_norm=0.5, lr=7e-4, lrschedule='linear', epsilon=1e-5, alpha=0.99, gamma=0.99, log_interval=100):
+    tf.reset_default_graph()
+    set_global_seeds(seed)
+
+    nenvs = env.num_envs
+    ob_space = env.observation_space
+    ac_space = env.action_space
+    num_procs = len(env.remotes) # HACK
+    model = Model(policy=policy, ob_space=ob_space, ac_space=ac_space, nenvs=nenvs, nsteps=nsteps, nstack=nstack, num_procs=num_procs, 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)
+    runner = Runner(env, model, nsteps=nsteps, nstack=nstack, gamma=gamma)
+
+    nbatch = nenvs*nsteps
+    tstart = time.time()
+    for update in range(1, total_timesteps//nbatch+1):
+        obs, states, rewards, masks, actions, values = runner.run()
+        policy_loss, value_loss, policy_entropy = model.train(obs, states, rewards, masks, actions, values)
+        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("value_loss", float(value_loss))
+            logger.record_tabular("explained_variance", float(ev))
+            logger.dump_tabular()
+    env.close()
+
+if __name__ == '__main__':
+    main()

baselines/a2c/policies.py

@@ -0,0 +1,120 @@
+import numpy as np
+import tensorflow as tf
+from baselines.a2c.utils import conv, fc, conv_to_fc, batch_to_seq, seq_to_batch, lstm, lnlstm, sample
+
+class LnLstmPolicy(object):
+    def __init__(self, sess, ob_space, ac_space, nenv, nsteps, nstack, nlstm=256, 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
+        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 = conv(tf.cast(X, tf.float32)/255., 'c1', nf=32, rf=8, stride=4, init_scale=np.sqrt(2))
+            h2 = conv(h, 'c2', nf=64, rf=4, stride=2, init_scale=np.sqrt(2))
+            h3 = conv(h2, 'c3', nf=64, rf=3, stride=1, init_scale=np.sqrt(2))
+            h3 = conv_to_fc(h3)
+            h4 = fc(h3, 'fc1', nh=512, init_scale=np.sqrt(2))
+            xs = batch_to_seq(h4, nenv, nsteps)
+            ms = batch_to_seq(M, nenv, nsteps)
+            h5, snew = lnlstm(xs, ms, S, 'lstm1', nh=nlstm)
+            h5 = seq_to_batch(h5)
+            pi = fc(h5, 'pi', nact, act=lambda x:x)
+            vf = fc(h5, 'v', 1, act=lambda x:x)
+
+        v0 = vf[:, 0]
+        a0 = sample(pi)
+        self.initial_state = np.zeros((nenv, nlstm*2), dtype=np.float32)
+
+        def step(ob, state, mask):
+            a, v, s = sess.run([a0, v0, snew], {X:ob, S:state, M:mask})
+            return a, v, s
+
+        def value(ob, state, mask):
+            return sess.run(v0, {X:ob, S:state, M:mask})
+
+        self.X = X
+        self.M = M
+        self.S = S
+        self.pi = pi
+        self.vf = vf
+        self.step = step
+        self.value = value
+
+class LstmPolicy(object):
+
+    def __init__(self, sess, ob_space, ac_space, nenv, nsteps, nstack, nlstm=256, 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
+        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 = conv(tf.cast(X, tf.float32)/255., 'c1', nf=32, rf=8, stride=4, init_scale=np.sqrt(2))
+            h2 = conv(h, 'c2', nf=64, rf=4, stride=2, init_scale=np.sqrt(2))
+            h3 = conv(h2, 'c3', nf=64, rf=3, stride=1, init_scale=np.sqrt(2))
+            h3 = conv_to_fc(h3)
+            h4 = fc(h3, 'fc1', nh=512, init_scale=np.sqrt(2))
+            xs = batch_to_seq(h4, nenv, nsteps)
+            ms = batch_to_seq(M, nenv, nsteps)
+            h5, snew = lstm(xs, ms, S, 'lstm1', nh=nlstm)
+            h5 = seq_to_batch(h5)
+            pi = fc(h5, 'pi', nact, act=lambda x:x)
+            vf = fc(h5, 'v', 1, act=lambda x:x)
+
+        v0 = vf[:, 0]
+        a0 = sample(pi)
+        self.initial_state = np.zeros((nenv, nlstm*2), dtype=np.float32)
+
+        def step(ob, state, mask):
+            a, v, s = sess.run([a0, v0, snew], {X:ob, S:state, M:mask})
+            return a, v, s
+
+        def value(ob, state, mask):
+            return sess.run(v0, {X:ob, S:state, M:mask})
+
+        self.X = X
+        self.M = M
+        self.S = S
+        self.pi = pi
+        self.vf = vf
+        self.step = step
+        self.value = value
+
+class CnnPolicy(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 = conv(tf.cast(X, tf.float32)/255., 'c1', nf=32, rf=8, stride=4, init_scale=np.sqrt(2))
+            h2 = conv(h, 'c2', nf=64, rf=4, stride=2, init_scale=np.sqrt(2))
+            h3 = conv(h2, 'c3', nf=64, rf=3, stride=1, init_scale=np.sqrt(2))
+            h3 = conv_to_fc(h3)
+            h4 = fc(h3, 'fc1', nh=512, init_scale=np.sqrt(2))
+            pi = fc(h4, 'pi', nact, act=lambda x:x)
+            vf = fc(h4, 'v', 1, act=lambda x:x)
+
+        v0 = vf[:, 0]
+        a0 = sample(pi)
+        self.initial_state = [] #not stateful
+
+        def step(ob, *_args, **_kwargs):
+            a, v = sess.run([a0, v0], {X:ob})
+            return a, v, [] #dummy state
+
+        def value(ob, *_args, **_kwargs):
+            return sess.run(v0, {X:ob})
+
+        self.X = X
+        self.pi = pi
+        self.vf = vf
+        self.step = step
+        self.value = value

baselines/a2c/run_atari.py

@@ -0,0 +1,45 @@
+#!/usr/bin/env python
+import os, logging, gym
+from baselines import logger
+from baselines.common import set_global_seeds
+from baselines import bench
+from baselines.a2c.a2c import learn
+from baselines.common.vec_env.subproc_vec_env import SubprocVecEnv
+from baselines.common.atari_wrappers import make_atari, wrap_deepmind
+from baselines.a2c.policies import CnnPolicy, LstmPolicy, LnLstmPolicy
+
+def train(env_id, num_timesteps, seed, policy, lrschedule, num_cpu):
+    def make_env(rank):
+        def _thunk():
+            env = make_atari(env_id)
+            env.seed(seed + rank)
+            env = bench.Monitor(env, logger.get_dir() and os.path.join(logger.get_dir(), str(rank)))
+            gym.logger.setLevel(logging.WARN)
+            return wrap_deepmind(env)
+        return _thunk
+    set_global_seeds(seed)
+    env = SubprocVecEnv([make_env(i) for i in range(num_cpu)])
+    if policy == 'cnn':
+        policy_fn = CnnPolicy
+    elif policy == 'lstm':
+        policy_fn = LstmPolicy
+    elif policy == 'lnlstm':
+        policy_fn = LnLstmPolicy
+    learn(policy_fn, env, seed, total_timesteps=int(num_timesteps * 1.1), lrschedule=lrschedule)
+    env.close()
+
+def main():
+    import argparse
+    parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter)
+    parser.add_argument('--env', help='environment ID', default='BreakoutNoFrameskip-v4')
+    parser.add_argument('--seed', help='RNG seed', type=int, default=0)
+    parser.add_argument('--policy', help='Policy architecture', choices=['cnn', 'lstm', 'lnlstm'], default='cnn')
+    parser.add_argument('--lrschedule', help='Learning rate schedule', choices=['constant', 'linear'], default='constant')
+    parser.add_argument('--num-timesteps', type=int, default=int(10e6))
+    args = parser.parse_args()
+    logger.configure()
+    train(args.env, num_timesteps=args.num_timesteps, seed=args.seed,
+        policy=args.policy, lrschedule=args.lrschedule, num_cpu=16)
+
+if __name__ == '__main__':
+    main()

baselines/a2c/utils.py

@@ -0,0 +1,255 @@
+import os
+import gym
+import numpy as np
+import tensorflow as tf
+from gym import spaces
+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, keep_dims=True)
+    ea0 = tf.exp(a0)
+    z0 = tf.reduce_sum(ea0, 1, keep_dims=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 mse(pred, target):
+    return tf.square(pred-target)/2.
+
+def ortho_init(scale=1.0):
+    def _ortho_init(shape, dtype, partition_info=None):
+        #lasagne ortho init for tf
+        shape = tuple(shape)
+        if len(shape) == 2:
+            flat_shape = shape
+        elif len(shape) == 4: # assumes NHWC
+            flat_shape = (np.prod(shape[:-1]), shape[-1])
+        else:
+            raise NotImplementedError
+        a = np.random.normal(0.0, 1.0, flat_shape)
+        u, _, v = np.linalg.svd(a, full_matrices=False)
+        q = u if u.shape == flat_shape else v # pick the one with the correct shape
+        q = q.reshape(shape)
+        return (scale * q[:shape[0], :shape[1]]).astype(np.float32)
+    return _ortho_init
+
+def conv(x, scope, nf, rf, stride, pad='VALID', act=tf.nn.relu, init_scale=1.0):
+    with tf.variable_scope(scope):
+        nin = x.get_shape()[3].value
+        w = tf.get_variable("w", [rf, rf, nin, nf], initializer=ortho_init(init_scale))
+        b = tf.get_variable("b", [nf], initializer=tf.constant_initializer(0.0))
+        z = tf.nn.conv2d(x, w, strides=[1, stride, stride, 1], padding=pad)+b
+        h = act(z)
+        return h
+
+def fc(x, scope, nh, act=tf.nn.relu, init_scale=1.0):
+    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(0.0))
+        z = tf.matmul(x, w)+b
+        h = act(z)
+        return h
+
+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()]
+    nsteps = len(xs)
+    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()]
+    nsteps = len(xs)
+    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 = []
+    r = 0
+    for reward, done in zip(rewards[::-1], dones[::-1]):
+        r = reward + gamma*r*(1.-done) # fixed off by one bug
+        discounted.append(r)
+    return discounted[::-1]
+
+def find_trainable_variables(key):
+    with tf.variable_scope(key):
+        return tf.trainable_variables()
+
+def make_path(f):
+    return os.makedirs(f, exist_ok=True)
+
+def constant(p):
+    return 1
+
+def linear(p):
+    return 1-p
+
+schedules = {
+    'linear':linear,
+    'constant':constant
+}
+
+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 myadd(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 my_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/acktr/README.md

@@ -0,0 +1,5 @@
+# ACKTR
+
+- Original paper: https://arxiv.org/abs/1708.05144
+- Baselines blog post: https://blog.openai.com/baselines-acktr-a2c/
+- `python -m baselines.acktr.run_atari` runs the algorithm for 40M frames = 10M timesteps on an Atari game. See help (`-h`) for more options.

baselines/acktr/acktr_cont.py

@@ -0,0 +1,142 @@
+import numpy as np
+import tensorflow as tf
+from baselines import logger
+from baselines import common
+from baselines.common import tf_util as U
+from baselines.acktr import kfac
+from baselines.acktr.filters import ZFilter
+
+def pathlength(path):
+    return path["reward"].shape[0]# Loss function that we'll differentiate to get the policy gradient
+
+def rollout(env, policy, max_pathlength, animate=False, obfilter=None):
+    """
+    Simulate the env and policy for max_pathlength steps
+    """
+    ob = env.reset()
+    prev_ob = np.float32(np.zeros(ob.shape))
+    if obfilter: ob = obfilter(ob)
+    terminated = False
+
+    obs = []
+    acs = []
+    ac_dists = []
+    logps = []
+    rewards = []
+    for _ in range(max_pathlength):
+        if animate:
+            env.render()
+        state = np.concatenate([ob, prev_ob], -1)
+        obs.append(state)
+        ac, ac_dist, logp = policy.act(state)
+        acs.append(ac)
+        ac_dists.append(ac_dist)
+        logps.append(logp)
+        prev_ob = np.copy(ob)
+        scaled_ac = env.action_space.low + (ac + 1.) * 0.5 * (env.action_space.high - env.action_space.low)
+        scaled_ac = np.clip(scaled_ac, env.action_space.low, env.action_space.high)
+        ob, rew, done, _ = env.step(scaled_ac)
+        if obfilter: ob = obfilter(ob)
+        rewards.append(rew)
+        if done:
+            terminated = True
+            break
+    return {"observation" : np.array(obs), "terminated" : terminated,
+            "reward" : np.array(rewards), "action" : np.array(acs),
+            "action_dist": np.array(ac_dists), "logp" : np.array(logps)}
+
+def learn(env, policy, vf, gamma, lam, timesteps_per_batch, num_timesteps,
+    animate=False, callback=None, desired_kl=0.002):
+
+    obfilter = ZFilter(env.observation_space.shape)
+
+    max_pathlength = env.spec.timestep_limit
+    stepsize = tf.Variable(initial_value=np.float32(np.array(0.03)), name='stepsize')
+    inputs, loss, loss_sampled = policy.update_info
+    optim = kfac.KfacOptimizer(learning_rate=stepsize, cold_lr=stepsize*(1-0.9), momentum=0.9, kfac_update=2,\
+                                epsilon=1e-2, stats_decay=0.99, async=1, cold_iter=1,
+                                weight_decay_dict=policy.wd_dict, max_grad_norm=None)
+    pi_var_list = []
+    for var in tf.trainable_variables():
+        if "pi" in var.name:
+            pi_var_list.append(var)
+
+    update_op, q_runner = optim.minimize(loss, loss_sampled, var_list=pi_var_list)
+    do_update = U.function(inputs, update_op)
+    U.initialize()
+
+    # start queue runners
+    enqueue_threads = []
+    coord = tf.train.Coordinator()
+    for qr in [q_runner, vf.q_runner]:
+        assert (qr != None)
+        enqueue_threads.extend(qr.create_threads(U.get_session(), coord=coord, start=True))
+
+    i = 0
+    timesteps_so_far = 0
+    while True:
+        if timesteps_so_far > num_timesteps:
+            break
+        logger.log("********** Iteration %i ************"%i)
+
+        # Collect paths until we have enough timesteps
+        timesteps_this_batch = 0
+        paths = []
+        while True:
+            path = rollout(env, policy, max_pathlength, animate=(len(paths)==0 and (i % 10 == 0) and animate), obfilter=obfilter)
+            paths.append(path)
+            n = pathlength(path)
+            timesteps_this_batch += n
+            timesteps_so_far += n
+            if timesteps_this_batch > timesteps_per_batch:
+                break
+
+        # Estimate advantage function
+        vtargs = []
+        advs = []
+        for path in paths:
+            rew_t = path["reward"]
+            return_t = common.discount(rew_t, gamma)
+            vtargs.append(return_t)
+            vpred_t = vf.predict(path)
+            vpred_t = np.append(vpred_t, 0.0 if path["terminated"] else vpred_t[-1])
+            delta_t = rew_t + gamma*vpred_t[1:] - vpred_t[:-1]
+            adv_t = common.discount(delta_t, gamma * lam)
+            advs.append(adv_t)
+        # Update value function
+        vf.fit(paths, vtargs)
+
+        # Build arrays for policy update
+        ob_no = np.concatenate([path["observation"] for path in paths])
+        action_na = np.concatenate([path["action"] for path in paths])
+        oldac_dist = np.concatenate([path["action_dist"] for path in paths])
+        adv_n = np.concatenate(advs)
+        standardized_adv_n = (adv_n - adv_n.mean()) / (adv_n.std() + 1e-8)
+
+        # Policy update
+        do_update(ob_no, action_na, standardized_adv_n)
+
+        min_stepsize = np.float32(1e-8)
+        max_stepsize = np.float32(1e0)
+        # Adjust stepsize
+        kl = policy.compute_kl(ob_no, oldac_dist)
+        if kl > desired_kl * 2:
+            logger.log("kl too high")
+            U.eval(tf.assign(stepsize, tf.maximum(min_stepsize, stepsize / 1.5)))
+        elif kl < desired_kl / 2:
+            logger.log("kl too low")
+            U.eval(tf.assign(stepsize, tf.minimum(max_stepsize, stepsize * 1.5)))            
+        else:
+            logger.log("kl just right!")
+
+        logger.record_tabular("EpRewMean", np.mean([path["reward"].sum() for path in paths]))
+        logger.record_tabular("EpRewSEM", np.std([path["reward"].sum()/np.sqrt(len(paths)) for path in paths]))
+        logger.record_tabular("EpLenMean", np.mean([pathlength(path) for path in paths]))
+        logger.record_tabular("KL", kl)
+        if callback:
+            callback()
+        logger.dump_tabular()
+        i += 1
+
+    coord.request_stop()
+    coord.join(enqueue_threads)

baselines/acktr/acktr_disc.py

@@ -0,0 +1,216 @@
+import os.path as osp
+import time
+import joblib
+import numpy as np
+import tensorflow as tf
+from baselines import logger
+
+from baselines.common import set_global_seeds, explained_variance
+
+from baselines.acktr.utils import discount_with_dones
+from baselines.acktr.utils import Scheduler, find_trainable_variables
+from baselines.acktr.utils import cat_entropy, mse
+from baselines.acktr import kfac
+
+
+class Model(object):
+
+    def __init__(self, policy, ob_space, ac_space, nenvs,total_timesteps, nprocs=32, nsteps=20,
+                 nstack=4, 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'):
+        config = tf.ConfigProto(allow_soft_placement=True,
+                                intra_op_parallelism_threads=nprocs,
+                                inter_op_parallelism_threads=nprocs)
+        config.gpu_options.allow_growth = True
+        self.sess = sess = tf.Session(config=config)
+        nact = ac_space.n
+        nbatch = nenvs * nsteps
+        A = tf.placeholder(tf.int32, [nbatch])
+        ADV = tf.placeholder(tf.float32, [nbatch])
+        R = tf.placeholder(tf.float32, [nbatch])
+        PG_LR = tf.placeholder(tf.float32, [])
+        VF_LR = tf.placeholder(tf.float32, [])
+
+        self.model = step_model = policy(sess, ob_space, ac_space, nenvs, 1, nstack, reuse=False)
+        self.model2 = train_model = policy(sess, ob_space, ac_space, nenvs, nsteps, nstack, reuse=True)
+
+        logpac = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=train_model.pi, labels=A)
+        self.logits = logits = train_model.pi
+
+        ##training loss
+        pg_loss = tf.reduce_mean(ADV*logpac)
+        entropy = tf.reduce_mean(cat_entropy(train_model.pi))
+        pg_loss = pg_loss - ent_coef * entropy
+        vf_loss = tf.reduce_mean(mse(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(logpac)
+        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("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, async=1, cold_iter=10, max_grad_norm=max_grad_norm)
+
+            update_stats_op = 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}
+            if states != []:
+                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
+
+        def save(save_path):
+            ps = sess.run(params)
+            joblib.dump(ps, save_path)
+
+        def load(load_path):
+            loaded_params = joblib.load(load_path)
+            restores = []
+            for p, loaded_p in zip(params, loaded_params):
+                restores.append(p.assign(loaded_p))
+            sess.run(restores)
+
+
+
+        self.train = train
+        self.save = save
+        self.load = load
+        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)
+
+class Runner(object):
+
+    def __init__(self, env, model, nsteps, nstack, gamma):
+        self.env = env
+        self.model = model
+        nh, nw, nc = env.observation_space.shape
+        nenv = env.num_envs
+        self.batch_ob_shape = (nenv*nsteps, nh, nw, nc*nstack)
+        self.obs = np.zeros((nenv, nh, nw, nc*nstack), dtype=np.uint8)
+        obs = env.reset()
+        self.update_obs(obs)
+        self.gamma = gamma
+        self.nsteps = nsteps
+        self.states = model.initial_state
+        self.dones = [False for _ in range(nenv)]
+
+    def update_obs(self, obs):
+        self.obs = np.roll(self.obs, shift=-1, axis=3)
+        self.obs[:, :, :, -1] = obs[:, :, :, 0]
+
+    def run(self):
+        mb_obs, mb_rewards, mb_actions, mb_values, mb_dones = [],[],[],[],[]
+        mb_states = self.states
+        for n in range(self.nsteps):
+            actions, values, states = self.model.step(self.obs, self.states, self.dones)
+            mb_obs.append(np.copy(self.obs))
+            mb_actions.append(actions)
+            mb_values.append(values)
+            mb_dones.append(self.dones)
+            obs, rewards, dones, _ = self.env.step(actions)
+            self.states = states
+            self.dones = dones
+            for n, done in enumerate(dones):
+                if done:
+                    self.obs[n] = self.obs[n]*0
+            self.update_obs(obs)
+            mb_rewards.append(rewards)
+        mb_dones.append(self.dones)
+        #batch of steps to batch of rollouts
+        mb_obs = np.asarray(mb_obs, dtype=np.uint8).swapaxes(1, 0).reshape(self.batch_ob_shape)
+        mb_rewards = np.asarray(mb_rewards, dtype=np.float32).swapaxes(1, 0)
+        mb_actions = np.asarray(mb_actions, dtype=np.int32).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]
+        mb_dones = mb_dones[:, 1:]
+        last_values = self.model.value(self.obs, self.states, self.dones).tolist()
+        #discount/bootstrap off value fn
+        for n, (rewards, dones, value) in enumerate(zip(mb_rewards, mb_dones, last_values)):
+            rewards = rewards.tolist()
+            dones = dones.tolist()
+            if dones[-1] == 0:
+                rewards = discount_with_dones(rewards+[value], dones+[0], self.gamma)[:-1]
+            else:
+                rewards = discount_with_dones(rewards, dones, self.gamma)
+            mb_rewards[n] = rewards
+        mb_rewards = mb_rewards.flatten()
+        mb_actions = mb_actions.flatten()
+        mb_values = mb_values.flatten()
+        mb_masks = mb_masks.flatten()
+        return mb_obs, mb_states, mb_rewards, mb_masks, mb_actions, mb_values
+
+def learn(policy, env, seed, total_timesteps=int(40e6), gamma=0.99, log_interval=1, nprocs=32, nsteps=20,
+                 nstack=4, 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'):
+    tf.reset_default_graph()
+    set_global_seeds(seed)
+
+    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, nstack=nstack, 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)
+    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()
+
+    runner = Runner(env, model, nsteps=nsteps, nstack=nstack, gamma=gamma)
+    nbatch = nenvs*nsteps
+    tstart = time.time()
+    coord = tf.train.Coordinator()
+    enqueue_threads = model.q_runner.create_threads(model.sess, coord=coord, start=True)
+    for update in range(1, total_timesteps//nbatch+1):
+        obs, states, rewards, masks, actions, values = runner.run()
+        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.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)
+    env.close()

baselines/acktr/filters.py

@@ -0,0 +1,98 @@
+from baselines.acktr.running_stat import RunningStat
+from collections import deque
+import numpy as np
+
+class Filter(object):
+    def __call__(self, x, update=True):
+        raise NotImplementedError
+    def reset(self):
+        pass
+
+class IdentityFilter(Filter):
+    def __call__(self, x, update=True):
+        return x
+
+class CompositionFilter(Filter):
+    def __init__(self, fs):
+        self.fs = fs
+    def __call__(self, x, update=True):
+        for f in self.fs:
+            x = f(x)
+        return x
+    def output_shape(self, input_space):
+        out = input_space.shape
+        for f in self.fs:
+            out = f.output_shape(out)
+        return out
+
+class ZFilter(Filter):
+    """
+    y = (x-mean)/std
+    using running estimates of mean,std
+    """
+
+    def __init__(self, shape, demean=True, destd=True, clip=10.0):
+        self.demean = demean
+        self.destd = destd
+        self.clip = clip
+
+        self.rs = RunningStat(shape)
+
+    def __call__(self, x, update=True):
+        if update: self.rs.push(x)
+        if self.demean:
+            x = x - self.rs.mean
+        if self.destd:
+            x = x / (self.rs.std+1e-8)
+        if self.clip:
+            x = np.clip(x, -self.clip, self.clip)
+        return x
+    def output_shape(self, input_space):
+        return input_space.shape
+
+class AddClock(Filter):
+    def __init__(self):
+        self.count = 0
+    def reset(self):
+        self.count = 0
+    def __call__(self, x, update=True):
+        return np.append(x, self.count/100.0)
+    def output_shape(self, input_space):
+        return (input_space.shape[0]+1,)
+
+class FlattenFilter(Filter):
+    def __call__(self, x, update=True):
+        return x.ravel()
+    def output_shape(self, input_space):
+        return (int(np.prod(input_space.shape)),)
+
+class Ind2OneHotFilter(Filter):
+    def __init__(self, n):
+        self.n = n
+    def __call__(self, x, update=True):
+        out = np.zeros(self.n)
+        out[x] = 1
+        return out
+    def output_shape(self, input_space):
+        return (input_space.n,)
+
+class DivFilter(Filter):
+    def __init__(self, divisor):
+        self.divisor = divisor
+    def __call__(self, x, update=True):
+        return x / self.divisor
+    def output_shape(self, input_space):
+        return input_space.shape
+
+class StackFilter(Filter):
+    def __init__(self, length):
+        self.stack = deque(maxlen=length)
+    def reset(self):
+        self.stack.clear()
+    def __call__(self, x, update=True):
+        self.stack.append(x)
+        while len(self.stack) < self.stack.maxlen:
+            self.stack.append(x)
+        return np.concatenate(self.stack, axis=-1)
+    def output_shape(self, input_space):
+        return input_space.shape[:-1] + (input_space.shape[-1] * self.stack.maxlen,)

baselines/acktr/kfac.py

@@ -0,0 +1,926 @@
+import tensorflow as tf
+import numpy as np
+import re
+from baselines.acktr.kfac_utils import *
+from functools import reduce
+
+KFAC_OPS = ['MatMul', 'Conv2D', 'BiasAdd']
+KFAC_DEBUG = False
+
+
+class KfacOptimizer():
+
+    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, 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 = 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 bewteen 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 +1,124 @@
+import tensorflow as tf
+import numpy as np
+
+
+def gmatmul(a, b, transpose_a=False, transpose_b=False, reduce_dim=None):
+    if reduce_dim == None:
+        # general batch matmul
+        if len(a.get_shape()) == 3 and len(b.get_shape()) == 3:
+            return tf.batch_matmul(a, b, adj_x=transpose_a, adj_y=transpose_b)
+        elif len(a.get_shape()) == 3 and len(b.get_shape()) == 2:
+            if transpose_b:
+                N = b.get_shape()[0].value
+            else:
+                N = b.get_shape()[1].value
+            B = a.get_shape()[0].value
+            if transpose_a:
+                K = a.get_shape()[1].value
+                a = tf.reshape(tf.transpose(a, [0, 2, 1]), [-1, K])
+            else:
+                K = a.get_shape()[-1].value
+                a = tf.reshape(a, [-1, K])
+            result = tf.matmul(a, b, transpose_b=transpose_b)
+            result = tf.reshape(result, [B, -1, N])
+            return result
+        elif len(a.get_shape()) == 2 and len(b.get_shape()) == 3:
+            if transpose_a:
+                M = a.get_shape()[1].value
+            else:
+                M = a.get_shape()[0].value
+            B = b.get_shape()[0].value
+            if transpose_b:
+                K = b.get_shape()[-1].value
+                b = tf.transpose(tf.reshape(b, [-1, K]), [1, 0])
+            else:
+                K = b.get_shape()[1].value
+                b = tf.transpose(tf.reshape(
+                    tf.transpose(b, [0, 2, 1]), [-1, K]), [1, 0])
+            result = tf.matmul(a, b, transpose_a=transpose_a)
+            result = tf.transpose(tf.reshape(result, [M, B, -1]), [1, 0, 2])
+            return result
+        else:
+            return tf.matmul(a, b, transpose_a=transpose_a, transpose_b=transpose_b)
+    else:
+        # 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/policies.py

@@ -0,0 +1,77 @@
+import numpy as np
+import tensorflow as tf
+from baselines.acktr.utils import conv, fc, dense, conv_to_fc, sample, kl_div
+import baselines.common.tf_util as U
+
+class CnnPolicy(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 = conv(tf.cast(X, tf.float32)/255., 'c1', nf=32, rf=8, stride=4, init_scale=np.sqrt(2))
+            h2 = conv(h, 'c2', nf=64, rf=4, stride=2, init_scale=np.sqrt(2))
+            h3 = conv(h2, 'c3', nf=32, rf=3, stride=1, init_scale=np.sqrt(2))
+            h3 = conv_to_fc(h3)
+            h4 = fc(h3, 'fc1', nh=512, init_scale=np.sqrt(2))
+            pi = fc(h4, 'pi', nact, act=lambda x:x)
+            vf = fc(h4, 'v', 1, act=lambda x:x)
+
+        v0 = vf[:, 0]
+        a0 = sample(pi)
+        self.initial_state = [] #not stateful
+
+        def step(ob, *_args, **_kwargs):
+            a, v = sess.run([a0, v0], {X:ob})
+            return a, v, [] #dummy state
+
+        def value(ob, *_args, **_kwargs):
+            return sess.run(v0, {X:ob})
+
+        self.X = X
+        self.pi = pi
+        self.vf = vf
+        self.step = step
+        self.value = value
+
+
+class GaussianMlpPolicy(object):
+    def __init__(self, ob_dim, ac_dim):
+        # Here we'll construct a bunch of expressions, which will be used in two places:
+        # (1) When sampling actions
+        # (2) When computing loss functions, for the policy update
+        # Variables specific to (1) have the word "sampled" in them,
+        # whereas variables specific to (2) have the word "old" in them
+        ob_no = tf.placeholder(tf.float32, shape=[None, ob_dim*2], name="ob") # batch of observations
+        oldac_na = tf.placeholder(tf.float32, shape=[None, ac_dim], name="ac") # batch of actions previous actions
+        oldac_dist = tf.placeholder(tf.float32, shape=[None, ac_dim*2], name="oldac_dist") # batch of actions previous action distributions
+        adv_n = tf.placeholder(tf.float32, shape=[None], name="adv") # advantage function estimate
+        wd_dict = {}
+        h1 = tf.nn.tanh(dense(ob_no, 64, "h1", weight_init=U.normc_initializer(1.0), bias_init=0.0, weight_loss_dict=wd_dict))
+        h2 = tf.nn.tanh(dense(h1, 64, "h2", weight_init=U.normc_initializer(1.0), bias_init=0.0, weight_loss_dict=wd_dict))
+        mean_na = dense(h2, ac_dim, "mean", weight_init=U.normc_initializer(0.1), bias_init=0.0, weight_loss_dict=wd_dict) # Mean control output
+        self.wd_dict = wd_dict
+        self.logstd_1a = logstd_1a = tf.get_variable("logstd", [ac_dim], tf.float32, tf.zeros_initializer()) # Variance on outputs
+        logstd_1a = tf.expand_dims(logstd_1a, 0)
+        std_1a = tf.exp(logstd_1a)
+        std_na = tf.tile(std_1a, [tf.shape(mean_na)[0], 1])
+        ac_dist = tf.concat([tf.reshape(mean_na, [-1, ac_dim]), tf.reshape(std_na, [-1, ac_dim])], 1)
+        sampled_ac_na = tf.random_normal(tf.shape(ac_dist[:,ac_dim:])) * ac_dist[:,ac_dim:] + ac_dist[:,:ac_dim] # This is the sampled action we'll perform.
+        logprobsampled_n = - U.sum(tf.log(ac_dist[:,ac_dim:]), axis=1) - 0.5 * tf.log(2.0*np.pi)*ac_dim - 0.5 * U.sum(tf.square(ac_dist[:,:ac_dim] - sampled_ac_na) / (tf.square(ac_dist[:,ac_dim:])), axis=1) # Logprob of sampled action
+        logprob_n = - U.sum(tf.log(ac_dist[:,ac_dim:]), axis=1) - 0.5 * tf.log(2.0*np.pi)*ac_dim - 0.5 * U.sum(tf.square(ac_dist[:,:ac_dim] - oldac_na) / (tf.square(ac_dist[:,ac_dim:])), axis=1) # Logprob of previous actions under CURRENT policy (whereas oldlogprob_n is under OLD policy)
+        kl = U.mean(kl_div(oldac_dist, ac_dist, ac_dim))
+        #kl = .5 * U.mean(tf.square(logprob_n - oldlogprob_n)) # Approximation of KL divergence between old policy used to generate actions, and new policy used to compute logprob_n
+        surr = - U.mean(adv_n * logprob_n) # Loss function that we'll differentiate to get the policy gradient
+        surr_sampled = - U.mean(logprob_n) # Sampled loss of the policy
+        self._act = U.function([ob_no], [sampled_ac_na, ac_dist, logprobsampled_n]) # Generate a new action and its logprob
+        #self.compute_kl = U.function([ob_no, oldac_na, oldlogprob_n], kl) # Compute (approximate) KL divergence between old policy and new policy
+        self.compute_kl = U.function([ob_no, oldac_dist], kl)
+        self.update_info = ((ob_no, oldac_na, adv_n), surr, surr_sampled) # Input and output variables needed for computing loss
+        U.initialize() # Initialize uninitialized TF variables
+
+    def act(self, ob):
+        ac, ac_dist, logp = self._act(ob[None])
+        return ac[0], ac_dist[0], logp[0]

baselines/acktr/run_atari.py

@@ -0,0 +1,38 @@
+#!/usr/bin/env python
+import os, logging, gym
+from baselines import logger
+from baselines.common import set_global_seeds
+from baselines import bench
+from baselines.acktr.acktr_disc import learn
+from baselines.common.vec_env.subproc_vec_env import SubprocVecEnv
+from baselines.common.atari_wrappers import make_atari, wrap_deepmind
+from baselines.acktr.policies import CnnPolicy
+
+def train(env_id, num_timesteps, seed, num_cpu):
+    def make_env(rank):
+        def _thunk():
+            env = make_atari(env_id)
+            env.seed(seed + rank)
+            env = bench.Monitor(env, logger.get_dir() and logger.get_dir())
+            gym.logger.setLevel(logging.WARN)
+            return wrap_deepmind(env)
+        return _thunk
+    set_global_seeds(seed)
+    env = SubprocVecEnv([make_env(i) for i in range(num_cpu)])
+    policy_fn = CnnPolicy
+    learn(policy_fn, env, seed, total_timesteps=int(num_timesteps * 1.1), nprocs=num_cpu)
+    env.close()
+
+def main():
+    import argparse
+    parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter)
+    parser.add_argument('--env', help='environment ID', default='BreakoutNoFrameskip-v4')
+    parser.add_argument('--seed', help='RNG seed', type=int, default=0)
+    parser.add_argument('--num-timesteps', type=int, default=int(10e6))
+    args = parser.parse_args()
+    logger.configure()
+    train(args.env, num_timesteps=args.num_timesteps, seed=args.seed, num_cpu=32)
+
+
+if __name__ == '__main__':
+    main()

baselines/acktr/run_mujoco.py

@@ -0,0 +1,43 @@
+#!/usr/bin/env python
+import argparse
+import logging
+import os
+import tensorflow as tf
+import gym
+from baselines import logger
+from baselines.common import set_global_seeds
+from baselines import bench
+from baselines.acktr.acktr_cont import learn
+from baselines.acktr.policies import GaussianMlpPolicy
+from baselines.acktr.value_functions import NeuralNetValueFunction
+
+def train(env_id, num_timesteps, seed):
+    env=gym.make(env_id)
+    env = bench.Monitor(env, logger.get_dir() and os.path.join(logger.get_dir(), str(rank)))
+    set_global_seeds(seed)
+    env.seed(seed)
+    gym.logger.setLevel(logging.WARN)
+
+    with tf.Session(config=tf.ConfigProto()):
+        ob_dim = env.observation_space.shape[0]
+        ac_dim = env.action_space.shape[0]
+        with tf.variable_scope("vf"):
+            vf = NeuralNetValueFunction(ob_dim, ac_dim)
+        with tf.variable_scope("pi"):
+            policy = GaussianMlpPolicy(ob_dim, ac_dim)
+
+        learn(env, policy=policy, vf=vf,
+            gamma=0.99, lam=0.97, timesteps_per_batch=2500,
+            desired_kl=0.002,
+            num_timesteps=num_timesteps, animate=False)
+
+        env.close()
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description='Run Mujoco benchmark.')
+    parser.add_argument('--seed', help='RNG seed', type=int, default=0)
+    parser.add_argument('--env', help='environment ID', type=str, default="Reacher-v1")
+    parser.add_argument('--num-timesteps', type=int, default=int(1e6))
+    args = parser.parse_args()
+    logger.configure()
+    train(args.env, num_timesteps=args.num_timesteps, seed=args.seed)

baselines/acktr/running_stat.py

@@ -0,0 +1,46 @@
+import numpy as np
+
+# http://www.johndcook.com/blog/standard_deviation/
+class RunningStat(object):
+    def __init__(self, shape):
+        self._n = 0
+        self._M = np.zeros(shape)
+        self._S = np.zeros(shape)
+    def push(self, x):
+        x = np.asarray(x)
+        assert x.shape == self._M.shape
+        self._n += 1
+        if self._n == 1:
+            self._M[...] = x
+        else:
+            oldM = self._M.copy()
+            self._M[...] = oldM + (x - oldM)/self._n
+            self._S[...] = self._S + (x - oldM)*(x - self._M)
+    @property
+    def n(self):
+        return self._n
+    @property
+    def mean(self):
+        return self._M
+    @property
+    def var(self):
+        return self._S/(self._n - 1) if self._n > 1 else np.square(self._M)
+    @property
+    def std(self):
+        return np.sqrt(self.var)
+    @property
+    def shape(self):
+        return self._M.shape
+
+def test_running_stat():
+    for shp in ((), (3,), (3,4)):
+        li = []
+        rs = RunningStat(shp)
+        for _ in range(5):
+            val = np.random.randn(*shp)
+            rs.push(val)
+            li.append(val)
+            m = np.mean(li, axis=0)
+            assert np.allclose(rs.mean, m)
+            v = np.square(m) if (len(li) == 1) else np.var(li, ddof=1, axis=0)
+            assert np.allclose(rs.var, v)

baselines/acktr/utils.py

@@ -0,0 +1,200 @@
+import os
+import numpy as np
+import tensorflow as tf
+import baselines.common.tf_util as U
+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 std(x):
+    mean = tf.reduce_mean(x)
+    var = tf.reduce_mean(tf.square(x-mean))
+    return tf.sqrt(var)
+
+def cat_entropy(logits):
+    a0 = logits - tf.reduce_max(logits, 1, keep_dims=True)
+    ea0 = tf.exp(a0)
+    z0 = tf.reduce_sum(ea0, 1, keep_dims=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 mse(pred, target):
+    return tf.square(pred-target)/2.
+
+def ortho_init(scale=1.0):
+    def _ortho_init(shape, dtype, partition_info=None):
+        #lasagne ortho init for tf
+        shape = tuple(shape)
+        if len(shape) == 2:
+            flat_shape = shape
+        elif len(shape) == 4: # assumes NHWC
+            flat_shape = (np.prod(shape[:-1]), shape[-1])
+        else:
+            raise NotImplementedError
+        a = np.random.normal(0.0, 1.0, flat_shape)
+        u, _, v = np.linalg.svd(a, full_matrices=False)
+        q = u if u.shape == flat_shape else v # pick the one with the correct shape
+        q = q.reshape(shape)
+        return (scale * q[:shape[0], :shape[1]]).astype(np.float32)
+    return _ortho_init
+
+def conv(x, scope, nf, rf, stride, pad='VALID', act=tf.nn.relu, init_scale=1.0):
+    with tf.variable_scope(scope):
+        nin = x.get_shape()[3].value
+        w = tf.get_variable("w", [rf, rf, nin, nf], initializer=ortho_init(init_scale))
+        b = tf.get_variable("b", [nf], initializer=tf.constant_initializer(0.0))
+        z = tf.nn.conv2d(x, w, strides=[1, stride, stride, 1], padding=pad)+b
+        h = act(z)
+        return h
+
+def fc(x, scope, nh, act=tf.nn.relu, init_scale=1.0):
+    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(0.0))
+        z = tf.matmul(x, w)+b
+        h = act(z)
+        return h
+
+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(U.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(U.scope_name().split('/')[0] + '_' + 'losses', weight_decay)
+
+        return tf.nn.bias_add(tf.matmul(x, w), b)
+
+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 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)
+
+def discount_with_dones(rewards, dones, gamma):
+    discounted = []
+    r = 0
+    for reward, done in zip(rewards[::-1], dones[::-1]):
+        r = reward + gamma*r*(1.-done) # fixed off by one bug
+        discounted.append(r)
+    return discounted[::-1]
+
+def find_trainable_variables(key):
+    with tf.variable_scope(key):
+        return tf.trainable_variables()
+
+def make_path(f):
+    return os.makedirs(f, exist_ok=True)
+
+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

baselines/acktr/value_functions.py

@@ -0,0 +1,50 @@
+from baselines import logger
+import numpy as np
+from baselines import common
+from baselines.common import tf_util as U
+import tensorflow as tf
+from baselines.acktr import kfac
+from baselines.acktr.utils import dense
+
+class NeuralNetValueFunction(object):
+    def __init__(self, ob_dim, ac_dim): #pylint: disable=W0613
+        X = tf.placeholder(tf.float32, shape=[None, ob_dim*2+ac_dim*2+2]) # batch of observations
+        vtarg_n = tf.placeholder(tf.float32, shape=[None], name='vtarg')
+        wd_dict = {}
+        h1 = tf.nn.elu(dense(X, 64, "h1", weight_init=U.normc_initializer(1.0), bias_init=0, weight_loss_dict=wd_dict))
+        h2 = tf.nn.elu(dense(h1, 64, "h2", weight_init=U.normc_initializer(1.0), bias_init=0, weight_loss_dict=wd_dict))
+        vpred_n = dense(h2, 1, "hfinal", weight_init=U.normc_initializer(1.0), bias_init=0, weight_loss_dict=wd_dict)[:,0]
+        sample_vpred_n = vpred_n + tf.random_normal(tf.shape(vpred_n))
+        wd_loss = tf.get_collection("vf_losses", None)
+        loss = U.mean(tf.square(vpred_n - vtarg_n)) + tf.add_n(wd_loss)
+        loss_sampled = U.mean(tf.square(vpred_n - tf.stop_gradient(sample_vpred_n)))
+        self._predict = U.function([X], vpred_n)
+        optim = kfac.KfacOptimizer(learning_rate=0.001, cold_lr=0.001*(1-0.9), momentum=0.9, \
+                                    clip_kl=0.3, epsilon=0.1, stats_decay=0.95, \
+                                    async=1, kfac_update=2, cold_iter=50, \
+                                    weight_decay_dict=wd_dict, max_grad_norm=None)
+        vf_var_list = []
+        for var in tf.trainable_variables():
+            if "vf" in var.name:
+                vf_var_list.append(var)
+
+        update_op, self.q_runner = optim.minimize(loss, loss_sampled, var_list=vf_var_list)
+        self.do_update = U.function([X, vtarg_n], update_op) #pylint: disable=E1101
+        U.initialize() # Initialize uninitialized TF variables
+    def _preproc(self, path):
+        l = pathlength(path)
+        al = np.arange(l).reshape(-1,1)/10.0
+        act = path["action_dist"].astype('float32')
+        X = np.concatenate([path['observation'], act, al, np.ones((l, 1))], axis=1)
+        return X
+    def predict(self, path):
+        return self._predict(self._preproc(path))
+    def fit(self, paths, targvals):
+        X = np.concatenate([self._preproc(p) for p in paths])
+        y = np.concatenate(targvals)
+        logger.record_tabular("EVBefore", common.explained_variance(self._predict(X), y))
+        for _ in range(25): self.do_update(X, y)
+        logger.record_tabular("EVAfter", common.explained_variance(self._predict(X), y))
+
+def pathlength(path):
+    return path["reward"].shape[0]

baselines/bench/__init__.py

@@ -1,3 +1,2 @@
 from baselines.bench.benchmarks import *
 from baselines.bench.monitor import *
-

baselines/bench/benchmarks.py

@@ -1,93 +1,132 @@
+import os.path as osp
+
 _atari7 = ['BeamRider', 'Breakout', 'Enduro', 'Pong', 'Qbert', 'Seaquest', 'SpaceInvaders']
 _atariexpl7 = ['Freeway', 'Gravitar', 'MontezumaRevenge', 'Pitfall', 'PrivateEye', 'Solaris', 'Venture']
 
 _BENCHMARKS = []
 
+
 def register_benchmark(benchmark):
     for b in _BENCHMARKS:
         if b['name'] == benchmark['name']:
-            raise ValueError('Benchmark with name %s already registered!'%b['name'])
+            raise ValueError('Benchmark with name %s already registered!' % b['name'])
     _BENCHMARKS.append(benchmark)
 
+
 def list_benchmarks():
     return [b['name'] for b in _BENCHMARKS]
 
+
 def get_benchmark(benchmark_name):
     for b in _BENCHMARKS:
         if b['name'] == benchmark_name:
             return b
     raise ValueError('%s not found! Known benchmarks: %s' % (benchmark_name, list_benchmarks()))
 
+
 def get_task(benchmark, env_id):
     """Get a task by env_id. Return None if the benchmark doesn't have the env"""
     return next(filter(lambda task: task['env_id'] == env_id, benchmark['tasks']), None)
 
+
+def find_task_for_env_id_in_any_benchmark(env_id):
+    for bm in _BENCHMARKS:
+        for task in bm["tasks"]:
+            if task["env_id"] == env_id:
+                return bm, task
+    return None, None
+
+
 _ATARI_SUFFIX = 'NoFrameskip-v4'
 
 register_benchmark({
-    'name' : 'Atari200M',
-    'description' :'7 Atari games from Mnih et al. (2013), with pixel observations, 200M frames',
-    'tasks'  : [{'env_id' : _game + _ATARI_SUFFIX, 'trials' : 2, 'num_timesteps' : int(200e6)} for _game in _atari7]
+    'name': 'Atari50M',
+    'description': '7 Atari games from Mnih et al. (2013), with pixel observations, 50M timesteps',
+    'tasks': [{'env_id': _game + _ATARI_SUFFIX, 'trials': 2, 'num_timesteps': int(50e6)} for _game in _atari7]
 })
 
 register_benchmark({
-    'name' : 'Atari40M',
-    'description' :'7 Atari games from Mnih et al. (2013), with pixel observations, 40M frames',
-    'tasks'  : [{'env_id' : _game + _ATARI_SUFFIX, 'trials' : 2, 'num_timesteps' : int(40e6)} for _game in _atari7]
+    'name': 'Atari10M',
+    'description': '7 Atari games from Mnih et al. (2013), with pixel observations, 10M timesteps',
+    'tasks': [{'env_id': _game + _ATARI_SUFFIX, 'trials': 2, 'num_timesteps': int(10e6)} for _game in _atari7]
 })
 
 register_benchmark({
-    'name' : 'Atari1Hr',
-    'description' :'7 Atari games from Mnih et al. (2013), with pixel observations, 1 hour of walltime',
-    'tasks'  : [{'env_id' : _game + _ATARI_SUFFIX, 'trials' : 2, 'num_seconds' : 60*60} for _game in _atari7]
+    'name': 'Atari1Hr',
+    'description': '7 Atari games from Mnih et al. (2013), with pixel observations, 1 hour of walltime',
+    'tasks': [{'env_id': _game + _ATARI_SUFFIX, 'trials': 2, 'num_seconds': 60 * 60} for _game in _atari7]
 })
 
 register_benchmark({
-    'name' : 'AtariExploration40M',
-    'description' :'7 Atari games emphasizing exploration, with pixel observations, 40M frames',
-    'tasks'  : [{'env_id' : _game + _ATARI_SUFFIX, 'trials' : 2, 'num_timesteps' : int(40e6)} for _game in _atariexpl7]
+    'name': 'AtariExploration10M',
+    'description': '7 Atari games emphasizing exploration, with pixel observations, 10M timesteps',
+    'tasks': [{'env_id': _game + _ATARI_SUFFIX, 'trials': 2, 'num_timesteps': int(10e6)} for _game in _atariexpl7]
 })
 
 
+
+
+# MuJoCo
+
 _mujocosmall = [
     'InvertedDoublePendulum-v1', 'InvertedPendulum-v1',
     'HalfCheetah-v1', 'Hopper-v1', 'Walker2d-v1',
     'Reacher-v1', 'Swimmer-v1']
-
 register_benchmark({
-    'name' : 'Mujoco1M',
-    'description' : 'Some small 2D MuJoCo tasks, run for 1M timesteps',
-    'tasks' : [{'env_id' : _envid, 'trials' : 3, 'num_timesteps' : int(1e6)} for _envid in _mujocosmall]
+    'name': 'Mujoco1M',
+    'description': 'Some small 2D MuJoCo tasks, run for 1M timesteps',
+    'tasks': [{'env_id': _envid, 'trials': 3, 'num_timesteps': int(1e6)} for _envid in _mujocosmall]
 })
-
-_roboschool_mujoco = [
-    'RoboschoolInvertedDoublePendulum-v0', 'RoboschoolInvertedPendulum-v0',      # cartpole
-    'RoboschoolHalfCheetah-v0', 'RoboschoolHopper-v0', 'RoboschoolWalker2d-v0',  # forward walkers
-    'RoboschoolReacher-v0'
+register_benchmark({
+    'name': 'MujocoWalkers',
+    'description': 'MuJoCo forward walkers, run for 8M, humanoid 100M',
+    'tasks': [
+        {'env_id': "Hopper-v1", 'trials': 4, 'num_timesteps': 8 * 1000000},
+        {'env_id': "Walker2d-v1", 'trials': 4, 'num_timesteps': 8 * 1000000},
+        {'env_id': "Humanoid-v1", 'trials': 4, 'num_timesteps': 100 * 1000000},
     ]
-
-register_benchmark({
-    'name' : 'RoboschoolMujoco2M',
-    'description' : 'Same small 2D tasks, still improving up to 2M',
-    'tasks' : [{'env_id' : _envid, 'trials' : 3, 'num_timesteps' : int(2e6)} for _envid in _roboschool_mujoco]
 })
 
+# Roboschool
 
-_atari50 =  [ # actually 49
-            'Alien', 'Amidar', 'Assault', 'Asterix', 'Asteroids', 
-            'Atlantis', 'BankHeist', 'BattleZone', 'BeamRider',  'Bowling', 
-            'Boxing', 'Breakout', 'Centipede', 'ChopperCommand', 'CrazyClimber', 
-            'DemonAttack', 'DoubleDunk',  'Enduro', 'FishingDerby', 'Freeway', 
-            'Frostbite', 'Gopher', 'Gravitar', 'IceHockey', 'Jamesbond',  
-            'Kangaroo', 'Krull', 'KungFuMaster', 'MontezumaRevenge', 'MsPacman', 
-            'NameThisGame', 'Pitfall', 'Pong', 'PrivateEye', 'Qbert', 
-            'Riverraid', 'RoadRunner', 'Robotank', 'Seaquest', 'SpaceInvaders', 
-            'StarGunner', 'Tennis', 'TimePilot', 'Tutankham', 'UpNDown', 
-            'Venture', 'VideoPinball', 'WizardOfWor', 'Zaxxon', 
+register_benchmark({
+    'name': 'Roboschool8M',
+    'description': 'Small 2D tasks, up to 30 minutes to complete on 8 cores',
+    'tasks': [
+        {'env_id': "RoboschoolReacher-v1", 'trials': 4, 'num_timesteps': 2 * 1000000},
+        {'env_id': "RoboschoolAnt-v1", 'trials': 4, 'num_timesteps': 8 * 1000000},
+        {'env_id': "RoboschoolHalfCheetah-v1", 'trials': 4, 'num_timesteps': 8 * 1000000},
+        {'env_id': "RoboschoolHopper-v1", 'trials': 4, 'num_timesteps': 8 * 1000000},
+        {'env_id': "RoboschoolWalker2d-v1", 'trials': 4, 'num_timesteps': 8 * 1000000},
+    ]
+})
+register_benchmark({
+    'name': 'RoboschoolHarder',
+    'description': 'Test your might!!! Up to 12 hours on 32 cores',
+    'tasks': [
+        {'env_id': "RoboschoolHumanoid-v1", 'trials': 4, 'num_timesteps': 100 * 1000000},
+        {'env_id': "RoboschoolHumanoidFlagrun-v1", 'trials': 4, 'num_timesteps': 200 * 1000000},
+        {'env_id': "RoboschoolHumanoidFlagrunHarder-v1", 'trials': 4, 'num_timesteps': 400 * 1000000},
+    ]
+})
+
+# Other
+
+_atari50 = [  # actually 47
+    'Alien', 'Amidar', 'Assault', 'Asterix', 'Asteroids',
+    'Atlantis', 'BankHeist', 'BattleZone', 'BeamRider', 'Bowling',
+    'Breakout', 'Centipede', 'ChopperCommand', 'CrazyClimber',
+    'DemonAttack', 'DoubleDunk', 'Enduro', 'FishingDerby', 'Freeway',
+    'Frostbite', 'Gopher', 'Gravitar', 'IceHockey', 'Jamesbond',
+    'Kangaroo', 'Krull', 'KungFuMaster', 'MontezumaRevenge', 'MsPacman',
+    'NameThisGame', 'Pitfall', 'Pong', 'PrivateEye', 'Qbert',
+    'RoadRunner', 'Robotank', 'Seaquest', 'SpaceInvaders', 'StarGunner',
+    'Tennis', 'TimePilot', 'Tutankham', 'UpNDown', 'Venture',
+    'VideoPinball', 'WizardOfWor', 'Zaxxon',
 ]
 
 register_benchmark({
-    'name' : 'Atari50_40M',
-    'description' :'7 Atari games from Mnih et al. (2013), with pixel observations, 40M frames',
-    'tasks'  : [{'env_id' : _game + _ATARI_SUFFIX, 'trials' : 3, 'num_timesteps' : int(40e6)} for _game in _atari50]
+    'name': 'Atari50_10M',
+    'description': '47 Atari games from Mnih et al. (2013), with pixel observations, 10M timesteps',
+    'tasks': [{'env_id': _game + _ATARI_SUFFIX, 'trials': 3, 'num_timesteps': int(10e6)} for _game in _atari50]
 })

baselines/bench/monitor.py

@@ -2,20 +2,17 @@ __all__ = ['Monitor', 'get_monitor_files', 'load_results']
 
 import gym
 from gym.core import Wrapper
-from os import path
 import time
 from glob import glob
-
-try:
-    import ujson as json # Not necessary for monitor writing, but very useful for monitor loading
-except ImportError:
-    import json
+import csv
+import os.path as osp
+import json
 
 class Monitor(Wrapper):
-    EXT = "monitor.json"
+    EXT = "monitor.csv"
     f = None
 
-    def __init__(self, env, filename, allow_early_resets=False):
+    def __init__(self, env, filename, allow_early_resets=False, reset_keywords=()):
         Wrapper.__init__(self, env=env)
         self.tstart = time.time()
         if filename is None:
@@ -23,50 +20,38 @@ class Monitor(Wrapper):
             self.logger = None
         else:
             if not filename.endswith(Monitor.EXT):
-                filename = filename + "." + Monitor.EXT
+                if osp.isdir(filename):
+                    filename = osp.join(filename, Monitor.EXT)
+                else:
+                    filename = filename + "." + Monitor.EXT
             self.f = open(filename, "wt")
-            self.logger = JSONLogger(self.f)
-            self.logger.writekvs({"t_start": self.tstart, "gym_version": gym.__version__,
-                "env_id": env.spec.id if env.spec else 'Unknown'})
+            self.f.write('#%s\n'%json.dumps({"t_start": self.tstart, "gym_version": gym.__version__,
+                "env_id": env.spec.id if env.spec else 'Unknown'}))
+            self.logger = csv.DictWriter(self.f, fieldnames=('r', 'l', 't')+reset_keywords)
+            self.logger.writeheader()
+
+        self.reset_keywords = reset_keywords
         self.allow_early_resets = allow_early_resets
         self.rewards = None
         self.needs_reset = True
         self.episode_rewards = []
         self.episode_lengths = []
         self.total_steps = 0
-        self.current_metadata = {} # extra info that gets injected into each log entry
-        # Useful for metalearning where we're modifying the environment externally
-        # But want our logs to know about these modifications
+        self.current_reset_info = {} # extra info about the current episode, that was passed in during reset()
 
-    def __getstate__(self): # XXX
-        d = self.__dict__.copy()
-        if self.f:
-            del d['f'], d['logger']
-            d['_filename'] = self.f.name
-            d['_num_episodes'] = len(self.episode_rewards)
-        else:
-            d['_filename'] = None
-        return d
-    def __setstate__(self, d):
-        filename = d.pop('_filename')
-        self.__dict__ = d
-        if filename is not None:
-            nlines = d.pop('_num_episodes') + 1
-            self.f = open(filename, "r+t")
-            for _ in range(nlines):
-                self.f.readline()
-            self.f.truncate()        
-            self.logger = JSONLogger(self.f)
-
-
-    def reset(self):
+    def _reset(self, **kwargs):
         if not self.allow_early_resets and not self.needs_reset:
             raise RuntimeError("Tried to reset an environment before done. If you want to allow early resets, wrap your env with Monitor(env, path, allow_early_resets=True)")
         self.rewards = []
         self.needs_reset = False
-        return self.env.reset()
+        for k in self.reset_keywords:
+            v = kwargs.get(k)
+            if v is None:
+                raise ValueError('Expected you to pass kwarg %s into reset'%k)
+            self.current_reset_info[k] = v
+        return self.env.reset(**kwargs)
 
-    def step(self, action):
+    def _step(self, action):
         if self.needs_reset:
             raise RuntimeError("Tried to step environment that needs reset")
         ob, rew, done, info = self.env.step(action)
@@ -75,10 +60,11 @@ class Monitor(Wrapper):
             self.needs_reset = True
             eprew = sum(self.rewards)
             eplen = len(self.rewards)
-            epinfo = {"r": eprew, "l": eplen, "t": round(time.time() - self.tstart, 6)}
-            epinfo.update(self.current_metadata)
+            epinfo = {"r": round(eprew, 6), "l": eplen, "t": round(time.time() - self.tstart, 6)}
+            epinfo.update(self.current_reset_info)
             if self.logger:
-                self.logger.writekvs(epinfo)
+                self.logger.writerow(epinfo)
+                self.f.flush()
             self.episode_rewards.append(eprew)
             self.episode_lengths.append(eplen)
             info['episode'] = epinfo
@@ -98,49 +84,40 @@ class Monitor(Wrapper):
     def get_episode_lengths(self):
         return self.episode_lengths
 
-class JSONLogger(object):
-    def __init__(self, file):
-        self.file = file
-
-    def writekvs(self, kvs):
-        for k,v in kvs.items():
-            if hasattr(v, 'dtype'):
-                v = v.tolist()
-                kvs[k] = float(v)
-        self.file.write(json.dumps(kvs) + '\n')
-        self.file.flush()
-
-
 class LoadMonitorResultsError(Exception):
     pass
 
 def get_monitor_files(dir):
-    return glob(path.join(dir, "*" + Monitor.EXT))
+    return glob(osp.join(dir, "*" + Monitor.EXT))
 
 def load_results(dir):
-    fnames = get_monitor_files(dir)
-    if not fnames:
+    import pandas
+    monitor_files = glob(osp.join(dir, "*monitor.*")) # get both csv and (old) json files
+    if not monitor_files:
         raise LoadMonitorResultsError("no monitor files of the form *%s found in %s" % (Monitor.EXT, dir))
-    episodes = []
+    dfs = []
     headers = []
-    for fname in fnames:
+    for fname in monitor_files:
         with open(fname, 'rt') as fh:
-            lines = fh.readlines()
-        header = json.loads(lines[0])
-        headers.append(header)
-        for line in lines[1:]:
-            episode = json.loads(line)
-            episode['abstime'] = header['t_start'] + episode['t']
-            del episode['t']
-            episodes.append(episode)
-    header0 = headers[0]
-    for header in headers[1:]:
-        assert header['env_id'] == header0['env_id'], "mixing data from two envs"
-    episodes = sorted(episodes, key=lambda e: e['abstime'])
-    return {
-        'env_info': {'env_id': header0['env_id'], 'gym_version': header0['gym_version']},
-        'episode_end_times': [e['abstime'] for e in episodes],
-        'episode_lengths': [e['l'] for e in episodes],
-        'episode_rewards': [e['r'] for e in episodes],
-        'initial_reset_time': min([min(header['t_start'] for header in headers)])
-    }
+            if fname.endswith('csv'):
+                firstline = fh.readline()
+                assert firstline[0] == '#'
+                header = json.loads(firstline[1:])
+                df = pandas.read_csv(fh, index_col=None)
+                headers.append(header)
+            elif fname.endswith('json'): # Deprecated json format
+                episodes = []
+                lines = fh.readlines()
+                header = json.loads(lines[0])
+                headers.append(header)
+                for line in lines[1:]:
+                    episode = json.loads(line)
+                    episodes.append(episode)
+                df = pandas.DataFrame(episodes)
+        df['t'] += header['t_start']
+        dfs.append(df)
+    df = pandas.concat(dfs)
+    df.sort_values('t', inplace=True)
+    df['t'] -= min(header['t_start'] for header in headers)
+    df.headers = headers # HACK to preserve backwards compatibility
+    return df

baselines/common/atari_wrappers.py

@@ -1,9 +1,8 @@
 import numpy as np
 from collections import deque
-from PIL import Image
 import gym
 from gym import spaces
-
+import cv2
 
 class NoopResetEnv(gym.Wrapper):
     def __init__(self, env, noop_max=30):
@@ -13,11 +12,16 @@ class NoopResetEnv(gym.Wrapper):
         gym.Wrapper.__init__(self, env)
         self.noop_max = noop_max
         self.override_num_noops = None
-        assert env.unwrapped.get_action_meanings()[0] == 'NOOP'
+        if isinstance(env.action_space, gym.spaces.MultiBinary):
+            self.noop_action = np.zeros(self.env.action_space.n, dtype=np.int64)
+        else:
+            # used for atari environments
+            self.noop_action = 0
+            assert env.unwrapped.get_action_meanings()[0] == 'NOOP'
 
-    def _reset(self):
+    def _reset(self, **kwargs):
         """ Do no-op action for a number of steps in [1, noop_max]."""
-        self.env.reset()
+        self.env.reset(**kwargs)
         if self.override_num_noops is not None:
             noops = self.override_num_noops
         else:
@@ -25,9 +29,9 @@ class NoopResetEnv(gym.Wrapper):
         assert noops > 0
         obs = None
         for _ in range(noops):
-            obs, _, done, _ = self.env.step(0)
+            obs, _, done, _ = self.env.step(self.noop_action)
             if done:
-                obs = self.env.reset()
+                obs = self.env.reset(**kwargs)
         return obs
 
 class FireResetEnv(gym.Wrapper):
@@ -37,14 +41,14 @@ class FireResetEnv(gym.Wrapper):
         assert env.unwrapped.get_action_meanings()[1] == 'FIRE'
         assert len(env.unwrapped.get_action_meanings()) >= 3
 
-    def _reset(self):
-        self.env.reset()
+    def _reset(self, **kwargs):
+        self.env.reset(**kwargs)
         obs, _, done, _ = self.env.step(1)
         if done:
-            self.env.reset()
+            self.env.reset(**kwargs)
         obs, _, done, _ = self.env.step(2)
         if done:
-            self.env.reset()
+            self.env.reset(**kwargs)
         return obs
 
 class EpisodicLifeEnv(gym.Wrapper):
@@ -70,13 +74,13 @@ class EpisodicLifeEnv(gym.Wrapper):
         self.lives = lives
         return obs, reward, done, info
 
-    def _reset(self):
+    def _reset(self, **kwargs):
         """Reset only when lives are exhausted.
         This way all states are still reachable even though lives are episodic,
         and the learner need not know about any of this behind-the-scenes.
         """
         if self.was_real_done:
-            obs = self.env.reset()
+            obs = self.env.reset(**kwargs)
         else:
             # no-op step to advance from terminal/lost life state
             obs, _, _, _ = self.env.step(0)
@@ -88,30 +92,26 @@ class MaxAndSkipEnv(gym.Wrapper):
         """Return only every `skip`-th frame"""
         gym.Wrapper.__init__(self, env)
         # most recent raw observations (for max pooling across time steps)
-        self._obs_buffer = deque(maxlen=2)
+        self._obs_buffer = np.zeros((2,)+env.observation_space.shape, dtype='uint8')
         self._skip       = skip
 
     def _step(self, action):
         """Repeat action, sum reward, and max over last observations."""
         total_reward = 0.0
         done = None
-        for _ in range(self._skip):
+        for i in range(self._skip):
             obs, reward, done, info = self.env.step(action)
-            self._obs_buffer.append(obs)
+            if i == self._skip - 2: self._obs_buffer[0] = obs
+            if i == self._skip - 1: self._obs_buffer[1] = obs
             total_reward += reward
             if done:
                 break
-        max_frame = np.max(np.stack(self._obs_buffer), axis=0)
+        # Note that the observation on the done=True frame
+        # doesn't matter
+        max_frame = self._obs_buffer.max(axis=0)
 
         return max_frame, total_reward, done, info
 
-    def _reset(self):
-        """Clear past frame buffer and init. to first obs. from inner env."""
-        self._obs_buffer.clear()
-        obs = self.env.reset()
-        self._obs_buffer.append(obs)
-        return obs
-
 class ClipRewardEnv(gym.RewardWrapper):
     def _reward(self, reward):
         """Bin reward to {+1, 0, -1} by its sign."""
@@ -121,52 +121,89 @@ class WarpFrame(gym.ObservationWrapper):
     def __init__(self, env):
         """Warp frames to 84x84 as done in the Nature paper and later work."""
         gym.ObservationWrapper.__init__(self, env)
-        self.res = 84
-        self.observation_space = spaces.Box(low=0, high=255, shape=(self.res, self.res, 1))
+        self.width = 84
+        self.height = 84
+        self.observation_space = spaces.Box(low=0, high=255, shape=(self.height, self.width, 1))
 
-    def _observation(self, obs):
-        frame = np.dot(obs.astype('float32'), np.array([0.299, 0.587, 0.114], 'float32'))
-        frame = np.array(Image.fromarray(frame).resize((self.res, self.res),
-            resample=Image.BILINEAR), dtype=np.uint8)
-        return frame.reshape((self.res, self.res, 1))
+    def _observation(self, frame):
+        frame = cv2.cvtColor(frame, cv2.COLOR_RGB2GRAY)
+        frame = cv2.resize(frame, (self.width, self.height), interpolation=cv2.INTER_AREA)
+        return frame[:, :, None]
 
 class FrameStack(gym.Wrapper):
     def __init__(self, env, k):
-        """Buffer observations and stack across channels (last axis)."""
+        """Stack k last frames.
+
+        Returns lazy array, which is much more memory efficient.
+
+        See Also
+        --------
+        baselines.common.atari_wrappers.LazyFrames
+        """
         gym.Wrapper.__init__(self, env)
         self.k = k
         self.frames = deque([], maxlen=k)
         shp = env.observation_space.shape
-        assert shp[2] == 1  # can only stack 1-channel frames
-        self.observation_space = spaces.Box(low=0, high=255, shape=(shp[0], shp[1], k))
+        self.observation_space = spaces.Box(low=0, high=255, shape=(shp[0], shp[1], shp[2] * k))
 
     def _reset(self):
-        """Clear buffer and re-fill by duplicating the first observation."""
         ob = self.env.reset()
-        for _ in range(self.k): self.frames.append(ob)
-        return self._observation()
+        for _ in range(self.k):
+            self.frames.append(ob)
+        return self._get_ob()
 
     def _step(self, action):
         ob, reward, done, info = self.env.step(action)
         self.frames.append(ob)
-        return self._observation(), reward, done, info
+        return self._get_ob(), reward, done, info
 
-    def _observation(self):
+    def _get_ob(self):
         assert len(self.frames) == self.k
-        return np.concatenate(self.frames, axis=2)
+        return LazyFrames(list(self.frames))
 
-def wrap_deepmind(env, episode_life=True, clip_rewards=True):
+class ScaledFloatFrame(gym.ObservationWrapper):
+    def _observation(self, observation):
+        # careful! This undoes the memory optimization, use
+        # with smaller replay buffers only.
+        return np.array(observation).astype(np.float32) / 255.0
+
+class LazyFrames(object):
+    def __init__(self, frames):
+        """This object ensures that common frames between the observations are only stored once.
+        It exists purely to optimize memory usage which can be huge for DQN's 1M frames replay
+        buffers.
+
+        This object should only be converted to numpy array before being passed to the model.
+
+        You'd not belive how complex the previous solution was."""
+        self._frames = frames
+
+    def __array__(self, dtype=None):
+        out = np.concatenate(self._frames, axis=2)
+        if dtype is not None:
+            out = out.astype(dtype)
+        return out
+
+def make_atari(env_id):
+    env = gym.make(env_id)
+    assert 'NoFrameskip' in env.spec.id
+    env = NoopResetEnv(env, noop_max=30)
+    env = MaxAndSkipEnv(env, skip=4)
+    return env
+
+def wrap_deepmind(env, episode_life=True, clip_rewards=True, frame_stack=False, scale=False):
     """Configure environment for DeepMind-style Atari.
-
-    Note: this does not include frame stacking!"""
-    assert 'NoFrameskip' in env.spec.id  # required for DeepMind-style skip
+    """
     if episode_life:
         env = EpisodicLifeEnv(env)
-    # env = NoopResetEnv(env, noop_max=30)
-    env = MaxAndSkipEnv(env, skip=4)
     if 'FIRE' in env.unwrapped.get_action_meanings():
         env = FireResetEnv(env)
     env = WarpFrame(env)
+    if scale:
+        env = ScaledFloatFrame(env)
     if clip_rewards:
         env = ClipRewardEnv(env)
+    if frame_stack:
+        env = FrameStack(env, 4)
     return env
+

baselines/common/atari_wrappers_deprecated.py

@@ -1,239 +0,0 @@
-import cv2
-import gym
-import numpy as np
-
-from collections import deque
-from gym import spaces
-
-
-class NoopResetEnv(gym.Wrapper):
-    def __init__(self, env=None, noop_max=30):
-        """Sample initial states by taking random number of no-ops on reset.
-        No-op is assumed to be action 0.
-        """
-        super(NoopResetEnv, self).__init__(env)
-        self.noop_max = noop_max
-        self.override_num_noops = None
-        assert env.unwrapped.get_action_meanings()[0] == 'NOOP'
-
-    def _reset(self):
-        """ Do no-op action for a number of steps in [1, noop_max]."""
-        self.env.reset()
-        if self.override_num_noops is not None:
-            noops = self.override_num_noops
-        else:
-            noops = np.random.randint(1, self.noop_max + 1)
-        assert noops > 0
-        obs = None
-        for _ in range(noops):
-            obs, _, done, _ = self.env.step(0)
-            if done:
-                obs = self.env.reset()
-        return obs
-
-
-class FireResetEnv(gym.Wrapper):
-    def __init__(self, env=None):
-        """For environments where the user need to press FIRE for the game to start."""
-        super(FireResetEnv, self).__init__(env)
-        assert env.unwrapped.get_action_meanings()[1] == 'FIRE'
-        assert len(env.unwrapped.get_action_meanings()) >= 3
-
-    def _reset(self):
-        self.env.reset()
-        obs, _, done, _ = self.env.step(1)
-        if done:
-            self.env.reset()
-        obs, _, done, _ = self.env.step(2)
-        if done:
-            self.env.reset()
-        return obs
-
-
-class EpisodicLifeEnv(gym.Wrapper):
-    def __init__(self, env=None):
-        """Make end-of-life == end-of-episode, but only reset on true game over.
-        Done by DeepMind for the DQN and co. since it helps value estimation.
-        """
-        super(EpisodicLifeEnv, self).__init__(env)
-        self.lives = 0
-        self.was_real_done = True
-        self.was_real_reset = False
-
-    def _step(self, action):
-        obs, reward, done, info = self.env.step(action)
-        self.was_real_done = done
-        # check current lives, make loss of life terminal,
-        # then update lives to handle bonus lives
-        lives = self.env.unwrapped.ale.lives()
-        if lives < self.lives and lives > 0:
-            # for Qbert somtimes we stay in lives == 0 condtion for a few frames
-            # so its important to keep lives > 0, so that we only reset once
-            # the environment advertises done.
-            done = True
-        self.lives = lives
-        return obs, reward, done, info
-
-    def _reset(self):
-        """Reset only when lives are exhausted.
-        This way all states are still reachable even though lives are episodic,
-        and the learner need not know about any of this behind-the-scenes.
-        """
-        if self.was_real_done:
-            obs = self.env.reset()
-            self.was_real_reset = True
-        else:
-            # no-op step to advance from terminal/lost life state
-            obs, _, _, _ = self.env.step(0)
-            self.was_real_reset = False
-        self.lives = self.env.unwrapped.ale.lives()
-        return obs
-
-
-class MaxAndSkipEnv(gym.Wrapper):
-    def __init__(self, env=None, skip=4):
-        """Return only every `skip`-th frame"""
-        super(MaxAndSkipEnv, self).__init__(env)
-        # most recent raw observations (for max pooling across time steps)
-        self._obs_buffer = deque(maxlen=2)
-        self._skip = skip
-
-    def _step(self, action):
-        total_reward = 0.0
-        done = None
-        for _ in range(self._skip):
-            obs, reward, done, info = self.env.step(action)
-            self._obs_buffer.append(obs)
-            total_reward += reward
-            if done:
-                break
-
-        max_frame = np.max(np.stack(self._obs_buffer), axis=0)
-
-        return max_frame, total_reward, done, info
-
-    def _reset(self):
-        """Clear past frame buffer and init. to first obs. from inner env."""
-        self._obs_buffer.clear()
-        obs = self.env.reset()
-        self._obs_buffer.append(obs)
-        return obs
-
-
-class ProcessFrame84(gym.ObservationWrapper):
-    def __init__(self, env=None):
-        super(ProcessFrame84, self).__init__(env)
-        self.observation_space = spaces.Box(low=0, high=255, shape=(84, 84, 1))
-
-    def _observation(self, obs):
-        return ProcessFrame84.process(obs)
-
-    @staticmethod
-    def process(frame):
-        if frame.size == 210 * 160 * 3:
-            img = np.reshape(frame, [210, 160, 3]).astype(np.float32)
-        elif frame.size == 250 * 160 * 3:
-            img = np.reshape(frame, [250, 160, 3]).astype(np.float32)
-        else:
-            assert False, "Unknown resolution."
-        img = img[:, :, 0] * 0.299 + img[:, :, 1] * 0.587 + img[:, :, 2] * 0.114
-        resized_screen = cv2.resize(img, (84, 110), interpolation=cv2.INTER_AREA)
-        x_t = resized_screen[18:102, :]
-        x_t = np.reshape(x_t, [84, 84, 1])
-        return x_t.astype(np.uint8)
-
-
-class ClippedRewardsWrapper(gym.RewardWrapper):
-    def _reward(self, reward):
-        """Change all the positive rewards to 1, negative to -1 and keep zero."""
-        return np.sign(reward)
-
-
-class LazyFrames(object):
-    def __init__(self, frames):
-        """This object ensures that common frames between the observations are only stored once.
-        It exists purely to optimize memory usage which can be huge for DQN's 1M frames replay
-        buffers.
-
-        This object should only be converted to numpy array before being passed to the model.
-
-        You'd not belive how complex the previous solution was."""
-        self._frames = frames
-
-    def __array__(self, dtype=None):
-        out = np.concatenate(self._frames, axis=2)
-        if dtype is not None:
-            out = out.astype(dtype)
-        return out
-
-
-class FrameStack(gym.Wrapper):
-    def __init__(self, env, k):
-        """Stack k last frames.
-
-        Returns lazy array, which is much more memory efficient.
-
-        See Also
-        --------
-        baselines.common.atari_wrappers.LazyFrames
-        """
-        gym.Wrapper.__init__(self, env)
-        self.k = k
-        self.frames = deque([], maxlen=k)
-        shp = env.observation_space.shape
-        self.observation_space = spaces.Box(low=0, high=255, shape=(shp[0], shp[1], shp[2] * k))
-
-    def _reset(self):
-        ob = self.env.reset()
-        for _ in range(self.k):
-            self.frames.append(ob)
-        return self._get_ob()
-
-    def _step(self, action):
-        ob, reward, done, info = self.env.step(action)
-        self.frames.append(ob)
-        return self._get_ob(), reward, done, info
-
-    def _get_ob(self):
-        assert len(self.frames) == self.k
-        return LazyFrames(list(self.frames))
-
-
-class ScaledFloatFrame(gym.ObservationWrapper):
-    def _observation(self, obs):
-        # careful! This undoes the memory optimization, use
-        # with smaller replay buffers only.
-        return np.array(obs).astype(np.float32) / 255.0
-
-
-def wrap_dqn(env):
-    """Apply a common set of wrappers for Atari games."""
-    assert 'NoFrameskip' in env.spec.id
-    env = EpisodicLifeEnv(env)
-    env = NoopResetEnv(env, noop_max=30)
-    env = MaxAndSkipEnv(env, skip=4)
-    if 'FIRE' in env.unwrapped.get_action_meanings():
-        env = FireResetEnv(env)
-    env = ProcessFrame84(env)
-    env = FrameStack(env, 4)
-    env = ClippedRewardsWrapper(env)
-    return env
-
-
-class A2cProcessFrame(gym.Wrapper):
-    def __init__(self, env):
-        gym.Wrapper.__init__(self, env)
-        self.observation_space = spaces.Box(low=0, high=255, shape=(84, 84, 1))
-
-    def _step(self, action):
-        ob, reward, done, info = self.env.step(action)
-        return A2cProcessFrame.process(ob), reward, done, info
-
-    def _reset(self):
-        return A2cProcessFrame.process(self.env.reset())
-
-    @staticmethod
-    def process(frame):
-        frame = cv2.cvtColor(frame, cv2.COLOR_RGB2GRAY)
-        frame = cv2.resize(frame, (84, 84), interpolation=cv2.INTER_AREA)
-        return frame.reshape(84, 84, 1)

baselines/common/azure_utils.py

@@ -3,14 +3,14 @@ import tempfile
 import zipfile
 
 from azure.common import AzureMissingResourceHttpError
-from azure.storage.blob import BlobService
+try:
+    from azure.storage.blob import BlobService
+except ImportError:
+    from azure.storage.blob import BlockBlobService as BlobService
 from shutil import unpack_archive
 from threading import Event
 
-"""TODOS:
-   - use Azure snapshots instead of hacky backups
-"""
-
+# TODOS: use Azure snapshots instead of hacky backups
 
 def fixed_list_blobs(service, *args, **kwargs):
     """By defualt list_containers only returns a subset of results.
@@ -34,7 +34,7 @@ def make_archive(source_path, dest_path):
     prefix_path = os.path.dirname(source_path)
     with zipfile.ZipFile(dest_path, "w", compression=zipfile.ZIP_STORED) as zf:
         if os.path.isdir(source_path):
-            for dirname, subdirs, files in os.walk(source_path):
+            for dirname, _subdirs, files in os.walk(source_path):
                 zf.write(dirname, os.path.relpath(dirname, prefix_path))
                 for filename in files:
                     filepath = os.path.join(dirname, filename)
@@ -114,18 +114,23 @@ class Container(object):
             arcpath = os.path.join(td, "archive.zip")
             for backup_blob_name in [blob_name, blob_name + '.backup']:
                 try:
-                    blob_size = self._service.get_blob_properties(
+                    properties = self._service.get_blob_properties(
                         blob_name=backup_blob_name,
                         container_name=self._container_name
-                    )['content-length']
+                    )
+                    if hasattr(properties, 'properties'):
+                        # Annoyingly, Azure has changed the API and this now returns a blob
+                        # instead of it's properties with up-to-date azure package.
+                        blob_size = properties.properties.content_length
+                    else:
+                        blob_size = properties['content-length']
                     if int(blob_size) > 0:
                         self._service.get_blob_to_path(
                             container_name=self._container_name,
                             blob_name=backup_blob_name,
                             file_path=arcpath,
                             max_connections=4,
-                            progress_callback=progress_callback,
-                            max_retries=10)
+                            progress_callback=progress_callback)
                         unpack_archive(arcpath, dest_path)
                         download_done.wait()
                         return True

baselines/common/distributions.py

@@ -2,7 +2,6 @@ import tensorflow as tf
 import numpy as np
 import baselines.common.tf_util as U
 from tensorflow.python.ops import math_ops
-from tensorflow.python.ops import nn
 
 class Pd(object):
     """
@@ -108,7 +107,7 @@ class BernoulliPdType(PdType):
 #     def flatparam(self):
 #         return self.logits
 #     def mode(self):
-#         return U.argmax(self.logits, axis=1)
+#         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):
@@ -118,7 +117,7 @@ class BernoulliPdType(PdType):
 #         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)
+#         return U.argmax(self.logits - tf.log(-tf.log(u)), axis=-1)
 
 class CategoricalPd(Pd):
     def __init__(self, logits):
@@ -126,27 +125,33 @@ class CategoricalPd(Pd):
     def flatparam(self):
         return self.logits
     def mode(self):
-        return U.argmax(self.logits, axis=1)
+        return U.argmax(self.logits, axis=-1)
     def neglogp(self, x):
-        return tf.nn.sparse_softmax_cross_entropy_with_logits(logits=self.logits, labels=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
+        #       the implementation does not allow second-order derivatives...
+        one_hot_actions = tf.one_hot(x, self.logits.get_shape().as_list()[-1])
+        return tf.nn.softmax_cross_entropy_with_logits(
+            logits=self.logits,
+            labels=one_hot_actions)
     def kl(self, other):
-        a0 = self.logits - U.max(self.logits, axis=1, keepdims=True)
-        a1 = other.logits - U.max(other.logits, axis=1, keepdims=True)
+        a0 = self.logits - U.max(self.logits, axis=-1, keepdims=True)
+        a1 = other.logits - U.max(other.logits, axis=-1, keepdims=True)
         ea0 = tf.exp(a0)
         ea1 = tf.exp(a1)
-        z0 = U.sum(ea0, axis=1, keepdims=True)
-        z1 = U.sum(ea1, axis=1, keepdims=True)
+        z0 = U.sum(ea0, axis=-1, keepdims=True)
+        z1 = U.sum(ea1, axis=-1, keepdims=True)
         p0 = ea0 / z0
-        return U.sum(p0 * (a0 - tf.log(z0) - a1 + tf.log(z1)), axis=1)
+        return U.sum(p0 * (a0 - tf.log(z0) - a1 + tf.log(z1)), axis=-1)
     def entropy(self):
-        a0 = self.logits - U.max(self.logits, axis=1, keepdims=True)
+        a0 = self.logits - U.max(self.logits, axis=-1, keepdims=True)
         ea0 = tf.exp(a0)
-        z0 = U.sum(ea0, axis=1, keepdims=True)
+        z0 = U.sum(ea0, axis=-1, keepdims=True)
         p0 = ea0 / z0
-        return U.sum(p0 * (tf.log(z0) - a0), axis=1)
+        return U.sum(p0 * (tf.log(z0) - a0), axis=-1)
     def sample(self):
         u = tf.random_uniform(tf.shape(self.logits))
-        return tf.argmax(self.logits - tf.log(-tf.log(u)), axis=1)
+        return tf.argmax(self.logits - tf.log(-tf.log(u)), axis=-1)
     @classmethod
     def fromflat(cls, flat):
         return cls(flat)
@@ -177,7 +182,7 @@ class MultiCategoricalPd(Pd):
 class DiagGaussianPd(Pd):
     def __init__(self, flat):
         self.flat = flat
-        mean, logstd = tf.split(axis=len(flat.get_shape()) - 1, num_or_size_splits=2, value=flat)
+        mean, logstd = tf.split(axis=len(flat.shape)-1, num_or_size_splits=2, value=flat)
         self.mean = mean
         self.logstd = logstd
         self.std = tf.exp(logstd)
@@ -186,14 +191,14 @@ class DiagGaussianPd(Pd):
     def mode(self):
         return self.mean
     def neglogp(self, x):
-        return 0.5 * U.sum(tf.square((x - self.mean) / self.std), axis=len(x.get_shape()) - 1) \
+        return 0.5 * U.sum(tf.square((x - self.mean) / self.std), axis=-1) \
                + 0.5 * np.log(2.0 * np.pi) * tf.to_float(tf.shape(x)[-1]) \
-               + U.sum(self.logstd, axis=len(x.get_shape()) - 1)
+               + U.sum(self.logstd, axis=-1)
     def kl(self, other):
         assert isinstance(other, DiagGaussianPd)
         return U.sum(other.logstd - self.logstd + (tf.square(self.std) + tf.square(self.mean - other.mean)) / (2.0 * tf.square(other.std)) - 0.5, axis=-1)
     def entropy(self):
-        return U.sum(self.logstd + .5 * np.log(2.0 * np.pi * np.e), -1)
+        return U.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))
     @classmethod
@@ -209,11 +214,11 @@ class BernoulliPd(Pd):
     def mode(self):
         return tf.round(self.ps)
     def neglogp(self, x):
-        return U.sum(tf.nn.sigmoid_cross_entropy_with_logits(logits=self.logits, labels=tf.to_float(x)), axis=1)
+        return U.sum(tf.nn.sigmoid_cross_entropy_with_logits(logits=self.logits, labels=tf.to_float(x)), axis=-1)
     def kl(self, other):
-        return U.sum(tf.nn.sigmoid_cross_entropy_with_logits(logits=other.logits, labels=self.ps), axis=1) - U.sum(tf.nn.sigmoid_cross_entropy_with_logits(logits=self.logits, labels=self.ps), axis=1)
+        return U.sum(tf.nn.sigmoid_cross_entropy_with_logits(logits=other.logits, labels=self.ps), axis=-1) - U.sum(tf.nn.sigmoid_cross_entropy_with_logits(logits=self.logits, labels=self.ps), axis=-1)
     def entropy(self):
-        return U.sum(tf.nn.sigmoid_cross_entropy_with_logits(logits=self.logits, labels=self.ps), axis=1)
+        return U.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))
@@ -286,4 +291,3 @@ def validate_probtype(probtype, pdparam):
     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
-

baselines/common/misc_util.py

@@ -4,7 +4,6 @@ import os
 import pickle
 import random
 import tempfile
-import time
 import zipfile
 
 
@@ -153,76 +152,6 @@ class RunningAvg(object):
         """Get the current estimate"""
         return self._value
 
-
-class SimpleMonitor(gym.Wrapper):
-    def __init__(self, env):
-        """Adds two qunatities to info returned by every step:
-
-            num_steps: int
-                Number of steps takes so far
-            rewards: [float]
-                All the cumulative rewards for the episodes completed so far.
-        """
-        super().__init__(env)
-        # current episode state
-        self._current_reward = None
-        self._num_steps = None
-        # temporary monitor state that we do not save
-        self._time_offset = None
-        self._total_steps = None
-        # monitor state
-        self._episode_rewards = []
-        self._episode_lengths = []
-        self._episode_end_times = []
-
-    def _reset(self):
-        obs = self.env.reset()
-        # recompute temporary state if needed
-        if self._time_offset is None:
-            self._time_offset = time.time()
-            if len(self._episode_end_times) > 0:
-                self._time_offset -= self._episode_end_times[-1]
-        if self._total_steps is None:
-            self._total_steps = sum(self._episode_lengths)
-        # update monitor state
-        if self._current_reward is not None:
-            self._episode_rewards.append(self._current_reward)
-            self._episode_lengths.append(self._num_steps)
-            self._episode_end_times.append(time.time() - self._time_offset)
-        # reset episode state
-        self._current_reward = 0
-        self._num_steps = 0
-
-        return obs
-
-    def _step(self, action):
-        obs, rew, done, info = self.env.step(action)
-        self._current_reward += rew
-        self._num_steps += 1
-        self._total_steps += 1
-        info['steps'] = self._total_steps
-        info['rewards'] = self._episode_rewards
-        return (obs, rew, done, info)
-
-    def get_state(self):
-        return {
-            'env_id': self.env.unwrapped.spec.id,
-            'episode_data': {
-                'episode_rewards': self._episode_rewards,
-                'episode_lengths': self._episode_lengths,
-                'episode_end_times': self._episode_end_times,
-                'initial_reset_time': 0,
-            }
-        }
-
-    def set_state(self, state):
-        assert state['env_id'] == self.env.unwrapped.spec.id
-        ed = state['episode_data']
-        self._episode_rewards = ed['episode_rewards']
-        self._episode_lengths = ed['episode_lengths']
-        self._episode_end_times = ed['episode_end_times']
-
-
 def boolean_flag(parser, name, default=False, help=None):
     """Add a boolean flag to argparse parser.
 
@@ -237,8 +166,9 @@ def boolean_flag(parser, name, default=False, help=None):
     help: str
         help string for the flag
     """
-    parser.add_argument("--" + name, action="store_true", default=default, help=help)
-    parser.add_argument("--no-" + name, action="store_false", dest=name)
+    dest = name.replace('-', '_')
+    parser.add_argument("--" + name, action="store_true", default=default, dest=dest, help=help)
+    parser.add_argument("--no-" + name, action="store_false", dest=dest)
 
 
 def get_wrapper_by_name(env, classname):

baselines/common/mpi_fork.py

@@ -1,6 +1,6 @@
 import os, subprocess, sys
 
-def mpi_fork(n):
+def mpi_fork(n, bind_to_core=False):
     """Re-launches the current script with workers
     Returns "parent" for original parent, "child" for MPI children
     """
@@ -13,7 +13,11 @@ def mpi_fork(n):
             OMP_NUM_THREADS="1",
             IN_MPI="1"
         )
-        subprocess.check_call(["mpirun", "-np", str(n), sys.executable] + sys.argv, env=env)
+        args = ["mpirun", "-np", str(n)]
+        if bind_to_core:
+            args += ["-bind-to", "core"]
+        args += [sys.executable] + sys.argv
+        subprocess.check_call(args, env=env)
         return "parent"
     else:
         return "child"

baselines/common/tf_util.py

@@ -6,51 +6,41 @@ import copy
 import os
 import collections
 
-
 # ================================================================
 # Make consistent with numpy
 # ================================================================
 
 clip = tf.clip_by_value
 
-
 def sum(x, axis=None, keepdims=False):
     axis = None if axis is None else [axis]
     return tf.reduce_sum(x, axis=axis, keep_dims=keepdims)
 
-
 def mean(x, axis=None, keepdims=False):
     axis = None if axis is None else [axis]
     return tf.reduce_mean(x, axis=axis, keep_dims=keepdims)
 
-
 def var(x, axis=None, keepdims=False):
     meanx = mean(x, axis=axis, keepdims=keepdims)
     return mean(tf.square(x - meanx), axis=axis, keepdims=keepdims)
 
-
 def std(x, axis=None, keepdims=False):
     return tf.sqrt(var(x, axis=axis, keepdims=keepdims))
 
-
 def max(x, axis=None, keepdims=False):
     axis = None if axis is None else [axis]
     return tf.reduce_max(x, axis=axis, keep_dims=keepdims)
 
-
 def min(x, axis=None, keepdims=False):
     axis = None if axis is None else [axis]
     return tf.reduce_min(x, axis=axis, keep_dims=keepdims)
 
-
 def concatenate(arrs, axis=0):
     return tf.concat(axis=axis, values=arrs)
 
-
 def argmax(x, axis=None):
     return tf.argmax(x, axis=axis)
 
-
 def switch(condition, then_expression, else_expression):
     """Switches between two operations depending on a scalar value (int or bool).
     Note that both `then_expression` and `else_expression`
@@ -72,35 +62,29 @@ def switch(condition, then_expression, else_expression):
 # Extras
 # ================================================================
 
-
 def l2loss(params):
     if len(params) == 0:
         return tf.constant(0.0)
     else:
         return tf.add_n([sum(tf.square(p)) for p in params])
 
-
 def lrelu(x, leak=0.2):
     f1 = 0.5 * (1 + leak)
     f2 = 0.5 * (1 - leak)
     return f1 * x + f2 * abs(x)
 
-
 def categorical_sample_logits(X):
     # https://github.com/tensorflow/tensorflow/issues/456
     U = tf.random_uniform(tf.shape(X))
     return argmax(X - tf.log(-tf.log(U)), axis=1)
 
-
 # ================================================================
 # Inputs
 # ================================================================
 
-
 def is_placeholder(x):
     return type(x) is tf.Tensor and len(x.op.inputs) == 0
 
-
 class TfInput(object):
     def __init__(self, name="(unnamed)"):
         """Generalized Tensorflow placeholder. The main differences are:
@@ -119,7 +103,6 @@ class TfInput(object):
         """Given data input it to the placeholder(s)."""
         raise NotImplemented()
 
-
 class PlacholderTfInput(TfInput):
     def __init__(self, placeholder):
         """Wrapper for regular tensorflow placeholder."""
@@ -132,7 +115,6 @@ class PlacholderTfInput(TfInput):
     def make_feed_dict(self, data):
         return {self._placeholder: data}
 
-
 class BatchInput(PlacholderTfInput):
     def __init__(self, shape, dtype=tf.float32, name=None):
         """Creates a placeholder for a batch of tensors of a given shape and dtype
@@ -148,7 +130,6 @@ class BatchInput(PlacholderTfInput):
         """
         super().__init__(tf.placeholder(dtype, [None] + list(shape), name=name))
 
-
 class Uint8Input(PlacholderTfInput):
     def __init__(self, shape, name=None):
         """Takes input in uint8 format which is cast to float32 and divided by 255
@@ -171,7 +152,6 @@ class Uint8Input(PlacholderTfInput):
     def get(self):
         return self._output
 
-
 def ensure_tf_input(thing):
     """Takes either tf.placeholder of TfInput and outputs equivalent TfInput"""
     if isinstance(thing, TfInput):
@@ -185,7 +165,6 @@ def ensure_tf_input(thing):
 # Mathematical utils
 # ================================================================
 
-
 def huber_loss(x, delta=1.0):
     """Reference: https://en.wikipedia.org/wiki/Huber_loss"""
     return tf.where(
@@ -198,7 +177,6 @@ def huber_loss(x, delta=1.0):
 # Optimizer utils
 # ================================================================
 
-
 def minimize_and_clip(optimizer, objective, var_list, clip_val=10):
     """Minimized `objective` using `optimizer` w.r.t. variables in
     `var_list` while ensure the norm of the gradients for each
@@ -210,7 +188,6 @@ def minimize_and_clip(optimizer, objective, var_list, clip_val=10):
             gradients[i] = (tf.clip_by_norm(grad, clip_val), var)
     return optimizer.apply_gradients(gradients)
 
-
 # ================================================================
 # Global session
 # ================================================================
@@ -219,7 +196,6 @@ def get_session():
     """Returns recently made Tensorflow session"""
     return tf.get_default_session()
 
-
 def make_session(num_cpu):
     """Returns a session that will use <num_cpu> CPU's only"""
     tf_config = tf.ConfigProto(
@@ -227,31 +203,25 @@ def make_session(num_cpu):
         intra_op_parallelism_threads=num_cpu)
     return tf.Session(config=tf_config)
 
-
 def single_threaded_session():
     """Returns a session which will only use a single CPU"""
     return make_session(1)
 
-
 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)
 
-
 def eval(expr, feed_dict=None):
     if feed_dict is None:
         feed_dict = {}
     return get_session().run(expr, feed_dict=feed_dict)
 
-
 VALUE_SETTERS = collections.OrderedDict()
 
-
 def set_value(v, val):
     global VALUE_SETTERS
     if v in VALUE_SETTERS:
@@ -262,17 +232,14 @@ def set_value(v, val):
         VALUE_SETTERS[v] = (set_op, set_endpoint)
     get_session().run(set_op, feed_dict={set_endpoint: val})
 
-
 # ================================================================
 # Saving variables
 # ================================================================
 
-
 def load_state(fname):
     saver = tf.train.Saver()
     saver.restore(get_session(), fname)
 
-
 def save_state(fname):
     os.makedirs(os.path.dirname(fname), exist_ok=True)
     saver = tf.train.Saver()
@@ -282,7 +249,6 @@ def save_state(fname):
 # Model components
 # ================================================================
 
-
 def normc_initializer(std=1.0):
     def _initializer(shape, dtype=None, partition_info=None):  # pylint: disable=W0613
         out = np.random.randn(*shape).astype(np.float32)
@@ -290,7 +256,6 @@ def normc_initializer(std=1.0):
         return tf.constant(out)
     return _initializer
 
-
 def conv2d(x, num_filters, name, filter_size=(3, 3), stride=(1, 1), pad="SAME", dtype=tf.float32, collections=None,
            summary_tag=None):
     with tf.variable_scope(name):
@@ -320,7 +285,6 @@ 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
 
-
 def dense(x, size, name, weight_init=None, bias=True):
     w = tf.get_variable(name + "/w", [x.get_shape()[1], size], initializer=weight_init)
     ret = tf.matmul(x, w)
@@ -330,7 +294,6 @@ def dense(x, size, name, weight_init=None, bias=True):
     else:
         return ret
 
-
 def wndense(x, size, name, init_scale=1.0):
     v = tf.get_variable(name + "/V", [int(x.get_shape()[1]), size],
                         initializer=tf.random_normal_initializer(0, 0.05))
@@ -342,11 +305,9 @@ def wndense(x, size, name, init_scale=1.0):
     scaler = g / tf.sqrt(sum(tf.square(v), axis=0, keepdims=True))
     return tf.reshape(scaler, [1, size]) * x + tf.reshape(b, [1, size])
 
-
 def densenobias(x, size, name, weight_init=None):
     return dense(x, size, name, weight_init=weight_init, bias=False)
 
-
 def dropout(x, pkeep, phase=None, mask=None):
     mask = tf.floor(pkeep + tf.random_uniform(tf.shape(x))) if mask is None else mask
     if phase is None:
@@ -354,13 +315,10 @@ def dropout(x, pkeep, phase=None, mask=None):
     else:
         return switch(phase, mask * x, pkeep * x)
 
-
 # ================================================================
 # 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
@@ -401,7 +359,6 @@ def function(inputs, outputs, updates=None, givens=None):
         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, check_nan=False):
         for inpt in inputs:
@@ -448,7 +405,6 @@ class _Function(object):
                 raise RuntimeError("Nan detected")
         return results
 
-
 def mem_friendly_function(nondata_inputs, data_inputs, outputs, batch_size):
     if isinstance(outputs, list):
         return _MemFriendlyFunction(nondata_inputs, data_inputs, outputs, batch_size)
@@ -456,7 +412,6 @@ def mem_friendly_function(nondata_inputs, data_inputs, outputs, batch_size):
         f = _MemFriendlyFunction(nondata_inputs, data_inputs, [outputs], batch_size)
         return lambda *inputs: f(*inputs)[0]
 
-
 class _MemFriendlyFunction(object):
     def __init__(self, nondata_inputs, data_inputs, outputs, batch_size):
         self.nondata_inputs = nondata_inputs
@@ -490,7 +445,6 @@ class _MemFriendlyFunction(object):
 # Modules
 # ================================================================
 
-
 class Module(object):
     def __init__(self, name):
         self.name = name
@@ -528,7 +482,6 @@ class Module(object):
         assert self.scope is not None, "need to call module once before getting variables"
         return tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, self.scope)
 
-
 def module(name):
     @functools.wraps
     def wrapper(f):
@@ -542,14 +495,11 @@ def module(name):
 # Graph traversal
 # ================================================================
 
-
 VARIABLES = {}
 
-
 def get_parents(node):
     return node.op.inputs
 
-
 def topsorted(outputs):
     """
     Topological sort via non-recursive depth-first search
@@ -586,7 +536,6 @@ def topsorted(outputs):
                 stack.append((j, 0))
     return out
 
-
 # ================================================================
 # Flat vectors
 # ================================================================
@@ -597,15 +546,12 @@ def var_shape(x):
         "shape function assumes that shape is fully known"
     return out
 
-
 def numel(x):
     return intprod(var_shape(x))
 
-
 def intprod(x):
     return int(np.prod(x))
 
-
 def flatgrad(loss, var_list, clip_norm=None):
     grads = tf.gradients(loss, var_list)
     if clip_norm is not None:
@@ -615,7 +561,6 @@ def flatgrad(loss, var_list, clip_norm=None):
         for (v, grad) in zip(var_list, grads)
     ])
 
-
 class SetFromFlat(object):
     def __init__(self, var_list, dtype=tf.float32):
         assigns = []
@@ -634,7 +579,6 @@ class SetFromFlat(object):
     def __call__(self, theta):
         get_session().run(self.op, feed_dict={self.theta: theta})
 
-
 class GetFlat(object):
     def __init__(self, var_list):
         self.op = tf.concat(axis=0, values=[tf.reshape(v, [numel(v)]) for v in var_list])
@@ -646,7 +590,6 @@ class GetFlat(object):
 # Misc
 # ================================================================
 
-
 def fancy_slice_2d(X, inds0, inds1):
     """
     like numpy X[inds0, inds1]
@@ -659,12 +602,10 @@ def fancy_slice_2d(X, inds0, inds1):
     Xflat = tf.reshape(X, [-1])
     return tf.gather(Xflat, inds0 * ncols + inds1)
 
-
 # ================================================================
 # Scopes
 # ================================================================
 
-
 def scope_vars(scope, trainable_only=False):
     """
     Get variables inside a scope
@@ -687,17 +628,14 @@ def scope_vars(scope, trainable_only=False):
         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 lengths_to_mask(lengths_b, max_length):
     """
     Turns a vector of lengths into a boolean mask
@@ -715,7 +653,6 @@ def lengths_to_mask(lengths_b, max_length):
     mask_bt = tf.expand_dims(tf.range(max_length), 0) < tf.expand_dims(lengths_b, 1)
     return mask_bt
 
-
 def in_session(f):
     @functools.wraps(f)
     def newfunc(*args, **kwargs):
@@ -723,10 +660,8 @@ def in_session(f):
             f(*args, **kwargs)
     return newfunc
 
-
 _PLACEHOLDER_CACHE = {}  # name -> (placeholder, dtype, shape)
 
-
 def get_placeholder(name, dtype, shape):
     if name in _PLACEHOLDER_CACHE:
         out, dtype1, shape1 = _PLACEHOLDER_CACHE[name]
@@ -737,15 +672,12 @@ def get_placeholder(name, dtype, shape):
         _PLACEHOLDER_CACHE[name] = (out, dtype, shape)
         return out
 
-
 def get_placeholder_cached(name):
     return _PLACEHOLDER_CACHE[name][0]
 
-
 def flattenallbut0(x):
     return tf.reshape(x, [-1, intprod(x.get_shape().as_list()[1:])])
 
-
 def reset():
     global _PLACEHOLDER_CACHE
     global VARIABLES

baselines/common/vec_env/__init__.py

@@ -0,0 +1,19 @@
+class VecEnv(object):
+    """
+    Vectorized environment base class
+    """
+    def step(self, vac):
+        """
+        Apply sequence of actions to sequence of environments
+        actions -> (observations, rewards, news)
+
+        where 'news' is a boolean vector indicating whether each element is new.
+        """
+        raise NotImplementedError
+    def reset(self):
+        """
+        Reset all environments
+        """
+        raise NotImplementedError
+    def close(self):
+        pass

baselines/common/vec_env/subproc_vec_env.py

@@ -0,0 +1,94 @@
+import numpy as np
+from multiprocessing import Process, Pipe
+from baselines.common.vec_env import VecEnv
+
+
+def worker(remote, parent_remote, env_fn_wrapper):
+    parent_remote.close()
+    env = env_fn_wrapper.x()
+    while True:
+        cmd, data = remote.recv()
+        if cmd == 'step':
+            ob, reward, done, info = env.step(data)
+            if done:
+                ob = env.reset()
+            remote.send((ob, reward, done, info))
+        elif cmd == 'reset':
+            ob = env.reset()
+            remote.send(ob)
+        elif cmd == 'reset_task':
+            ob = env.reset_task()
+            remote.send(ob)
+        elif cmd == 'close':
+            remote.close()
+            break
+        elif cmd == 'get_spaces':
+            remote.send((env.action_space, env.observation_space))
+        else:
+            raise NotImplementedError
+
+
+class CloudpickleWrapper(object):
+    """
+    Uses cloudpickle to serialize contents (otherwise multiprocessing tries to use pickle)
+    """
+    def __init__(self, x):
+        self.x = x
+    def __getstate__(self):
+        import cloudpickle
+        return cloudpickle.dumps(self.x)
+    def __setstate__(self, ob):
+        import pickle
+        self.x = pickle.loads(ob)
+
+
+class SubprocVecEnv(VecEnv):
+    def __init__(self, env_fns):
+        """
+        envs: list of gym environments to run in subprocesses
+        """
+        self.closed = False
+        nenvs = len(env_fns)
+        self.remotes, self.work_remotes = zip(*[Pipe() for _ in range(nenvs)])
+        self.ps = [Process(target=worker, args=(work_remote, remote, CloudpickleWrapper(env_fn)))
+            for (work_remote, remote, env_fn) in zip(self.work_remotes, self.remotes, env_fns)]
+        for p in self.ps:
+            p.daemon = True # if the main process crashes, we should not cause things to hang
+            p.start()
+        for remote in self.work_remotes:
+            remote.close()
+
+        self.remotes[0].send(('get_spaces', None))
+        self.action_space, self.observation_space = self.remotes[0].recv()
+
+
+    def step(self, actions):
+        for remote, action in zip(self.remotes, actions):
+            remote.send(('step', action))
+        results = [remote.recv() for remote in self.remotes]
+        obs, rews, dones, infos = zip(*results)
+        return np.stack(obs), np.stack(rews), np.stack(dones), infos
+
+    def reset(self):
+        for remote in self.remotes:
+            remote.send(('reset', None))
+        return np.stack([remote.recv() for remote in self.remotes])
+
+    def reset_task(self):
+        for remote in self.remotes:
+            remote.send(('reset_task', None))
+        return np.stack([remote.recv() for remote in self.remotes])
+
+    def close(self):
+        if self.closed:
+            return
+
+        for remote in self.remotes:
+            remote.send(('close', None))
+        for p in self.ps:
+            p.join()
+        self.closed = True
+
+    @property
+    def num_envs(self):
+        return len(self.remotes)

baselines/ddpg/README.md

@@ -0,0 +1,5 @@
+# DDPG
+
+- Original paper: https://arxiv.org/abs/1509.02971
+- Baselines post: https://blog.openai.com/better-exploration-with-parameter-noise/
+- `python -m baselines.ddpg.main` runs the algorithm for 1M frames = 10M timesteps on a Mujoco environment. See help (`-h`) for more options.

baselines/ddpg/ddpg.py

@@ -0,0 +1,372 @@
+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.common.mpi_adam import MpiAdam
+import baselines.common.tf_util as U
+from baselines.common.mpi_running_mean_std import RunningMeanStd
+from baselines.ddpg.util import reduce_std, mpi_mean
+
+
+def normalize(x, stats):
+    if stats is None:
+        return x
+    return (x - stats.mean) / stats.std
+
+
+def denormalize(x, stats):
+    if stats is None:
+        return x
+    return x * stats.std + stats.mean
+
+
+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)
+
+
+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:
+            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:
+            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(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),
+        adaptive_param_noise=True, adaptive_param_noise_policy_threshold=.1,
+        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
+        self.tau = tau
+        self.memory = memory
+        self.normalize_observations = normalize_observations
+        self.normalize_returns = normalize_returns
+        self.action_noise = action_noise
+        self.param_noise = param_noise
+        self.action_range = action_range
+        self.return_range = return_range
+        self.observation_range = observation_range
+        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
+
+        # Observation normalization.
+        if self.normalize_observations:
+            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.variable_scope('ret_rms'):
+                self.ret_rms = RunningMeanStd()
+        else:
+            self.ret_rms = None
+
+        # Create target networks.
+        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(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()
+
+    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, normalized_obs0):
+        assert self.param_noise is not None
+
+        # Configure perturbed actor.
+        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.
+        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 'kernel' in var.name 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
+            assert 'kernel' in M.name
+            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
+
+    def pi(self, obs, apply_noise=True, compute_Q=True):
+        if self.param_noise is not None and apply_noise:
+            actor_tf = self.perturbed_actor_tf
+        else:
+            actor_tf = self.actor_tf
+        feed_dict = {self.obs0: [obs]}
+        if compute_Q:
+            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
+        action = action.flatten()
+        if self.action_noise is not None and apply_noise:
+            noise = self.action_noise()
+            assert noise.shape == action.shape
+            action += noise
+        action = np.clip(action, self.action_range[0], self.action_range[1])
+        return action, q
+
+    def store_transition(self, obs0, action, reward, obs1, terminal1):
+        reward *= self.reward_scale
+        self.memory.append(obs0, action, reward, obs1, terminal1)
+        if self.normalize_observations:
+            self.obs_rms.update(np.array([obs0]))
+
+    def train(self):
+        # Get a batch.
+        batch = self.memory.sample(batch_size=self.batch_size)
+
+        if self.normalize_returns and self.enable_popart:
+            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())
+            self.sess.run(self.renormalize_Q_outputs_op, feed_dict={
+                self.old_std : np.array([old_std]),
+                self.old_mean : np.array([old_mean]),
+            })
+
+            # 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:
+            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
+
+    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)
+
+    def update_target_net(self):
+        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)
+        values = self.sess.run(self.stats_ops, feed_dict={
+            self.obs0: self.stats_sample['obs0'],
+            self.actions: self.stats_sample['actions'],
+        })
+
+        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()}
+
+        return stats
+
+    def adapt_param_noise(self):
+        if self.param_noise is None:
+            return 0.
+        
+        # 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,
+        })
+
+        mean_distance = mpi_mean(distance)
+        self.param_noise.adapt(mean_distance)
+        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:
+            self.sess.run(self.perturb_policy_ops, feed_dict={
+                self.param_noise_stddev: self.param_noise.current_stddev,
+            })

baselines/ddpg/main.py

@@ -0,0 +1,124 @@
+import argparse
+import time
+import os
+import logging
+from baselines import logger, bench
+from baselines.common.misc_util import (
+    set_global_seeds,
+    boolean_flag,
+)
+import baselines.ddpg.training as training
+from baselines.ddpg.models import Actor, Critic
+from baselines.ddpg.memory import Memory
+from baselines.ddpg.noise import *
+
+import gym
+import tensorflow as tf
+from mpi4py import MPI
+
+def run(env_id, seed, noise_type, layer_norm, evaluation, **kwargs):
+    # Configure things.
+    rank = MPI.COMM_WORLD.Get_rank()
+    if rank != 0:
+        logger.set_level(logger.DISABLED)
+
+    # Create envs.
+    env = gym.make(env_id)
+    env = bench.Monitor(env, logger.get_dir() and os.path.join(logger.get_dir(), str(rank)))
+    gym.logger.setLevel(logging.WARN)
+
+    if evaluation and rank==0:
+        eval_env = gym.make(env_id)
+        eval_env = bench.Monitor(eval_env, os.path.join(logger.get_dir(), 'gym_eval'))
+        env = bench.Monitor(env, None)
+    else:
+        eval_env = None
+
+    # Parse noise_type
+    action_noise = None
+    param_noise = None
+    nb_actions = env.action_space.shape[-1]
+    for current_noise_type in noise_type.split(','):
+        current_noise_type = current_noise_type.strip()
+        if current_noise_type == 'none':
+            pass
+        elif 'adaptive-param' in current_noise_type:
+            _, stddev = current_noise_type.split('_')
+            param_noise = AdaptiveParamNoiseSpec(initial_stddev=float(stddev), desired_action_stddev=float(stddev))
+        elif 'normal' in current_noise_type:
+            _, stddev = current_noise_type.split('_')
+            action_noise = NormalActionNoise(mu=np.zeros(nb_actions), sigma=float(stddev) * np.ones(nb_actions))
+        elif 'ou' in current_noise_type:
+            _, stddev = current_noise_type.split('_')
+            action_noise = OrnsteinUhlenbeckActionNoise(mu=np.zeros(nb_actions), sigma=float(stddev) * np.ones(nb_actions))
+        else:
+            raise RuntimeError('unknown noise type "{}"'.format(current_noise_type))
+
+    # Configure components.
+    memory = Memory(limit=int(1e6), action_shape=env.action_space.shape, observation_shape=env.observation_space.shape)
+    critic = Critic(layer_norm=layer_norm)
+    actor = Actor(nb_actions, layer_norm=layer_norm)
+
+    # Seed everything to make things reproducible.
+    seed = seed + 1000000 * rank
+    logger.info('rank {}: seed={}, logdir={}'.format(rank, seed, logger.get_dir()))
+    tf.reset_default_graph()
+    set_global_seeds(seed)
+    env.seed(seed)
+    if eval_env is not None:
+        eval_env.seed(seed)
+
+    # Disable logging for rank != 0 to avoid noise.
+    if rank == 0:
+        start_time = time.time()
+    training.train(env=env, eval_env=eval_env, param_noise=param_noise,
+        action_noise=action_noise, actor=actor, critic=critic, memory=memory, **kwargs)
+    env.close()
+    if eval_env is not None:
+        eval_env.close()
+    if rank == 0:
+        logger.info('total runtime: {}s'.format(time.time() - start_time))
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter)
+
+    parser.add_argument('--env-id', type=str, default='HalfCheetah-v1')
+    boolean_flag(parser, 'render-eval', default=False)
+    boolean_flag(parser, 'layer-norm', default=True)
+    boolean_flag(parser, 'render', default=False)
+    boolean_flag(parser, 'normalize-returns', default=False)
+    boolean_flag(parser, 'normalize-observations', default=True)
+    parser.add_argument('--seed', help='RNG seed', type=int, default=0)
+    parser.add_argument('--critic-l2-reg', type=float, default=1e-2)
+    parser.add_argument('--batch-size', type=int, default=64)  # per MPI worker
+    parser.add_argument('--actor-lr', type=float, default=1e-4)
+    parser.add_argument('--critic-lr', type=float, default=1e-3)
+    boolean_flag(parser, 'popart', default=False)
+    parser.add_argument('--gamma', type=float, default=0.99)
+    parser.add_argument('--reward-scale', type=float, default=1.)
+    parser.add_argument('--clip-norm', type=float, default=None)
+    parser.add_argument('--nb-epochs', type=int, default=500)  # with default settings, perform 1M steps total
+    parser.add_argument('--nb-epoch-cycles', type=int, default=20)
+    parser.add_argument('--nb-train-steps', type=int, default=50)  # per epoch cycle and MPI worker
+    parser.add_argument('--nb-eval-steps', type=int, default=100)  # per epoch cycle and MPI worker
+    parser.add_argument('--nb-rollout-steps', type=int, default=100)  # per epoch cycle and MPI worker
+    parser.add_argument('--noise-type', type=str, default='adaptive-param_0.2')  # choices are adaptive-param_xx, ou_xx, normal_xx, none
+    parser.add_argument('--num-timesteps', type=int, default=None)
+    boolean_flag(parser, 'evaluation', default=False)
+    args = parser.parse_args()
+    # we don't directly specify timesteps for this script, so make sure that if we do specify them
+    # they agree with the other parameters
+    if args.num_timesteps is not None:
+        assert(args.num_timesteps == args.nb_epochs * args.nb_epoch_cycles * args.nb_rollout_steps)
+    dict_args = vars(args)
+    del dict_args['num_timesteps']
+    return dict_args
+
+
+if __name__ == '__main__':
+    args = parse_args()
+    if MPI.COMM_WORLD.Get_rank() == 0:
+        logger.configure()
+    # Run actual script.
+    run(**args)

baselines/ddpg/memory.py

@@ -0,0 +1,83 @@
+import numpy as np
+
+
+class RingBuffer(object):
+    def __init__(self, maxlen, shape, dtype='float32'):
+        self.maxlen = maxlen
+        self.start = 0
+        self.length = 0
+        self.data = np.zeros((maxlen,) + shape).astype(dtype)
+
+    def __len__(self):
+        return self.length
+
+    def __getitem__(self, idx):
+        if idx < 0 or idx >= self.length:
+            raise KeyError()
+        return self.data[(self.start + idx) % self.maxlen]
+
+    def get_batch(self, idxs):
+        return self.data[(self.start + idxs) % self.maxlen]
+
+    def append(self, v):
+        if self.length < self.maxlen:
+            # We have space, simply increase the length.
+            self.length += 1
+        elif self.length == self.maxlen:
+            # No space, "remove" the first item.
+            self.start = (self.start + 1) % self.maxlen
+        else:
+            # This should never happen.
+            raise RuntimeError()
+        self.data[(self.start + self.length - 1) % self.maxlen] = v
+
+
+def array_min2d(x):
+    x = np.array(x)
+    if x.ndim >= 2:
+        return x
+    return x.reshape(-1, 1)
+
+
+class Memory(object):
+    def __init__(self, limit, action_shape, observation_shape):
+        self.limit = limit
+
+        self.observations0 = RingBuffer(limit, shape=observation_shape)
+        self.actions = RingBuffer(limit, shape=action_shape)
+        self.rewards = RingBuffer(limit, shape=(1,))
+        self.terminals1 = RingBuffer(limit, shape=(1,))
+        self.observations1 = RingBuffer(limit, shape=observation_shape)
+
+    def sample(self, batch_size):
+        # Draw such that we always have a proceeding element.
+        batch_idxs = np.random.random_integers(self.nb_entries - 2, size=batch_size)
+
+        obs0_batch = self.observations0.get_batch(batch_idxs)
+        obs1_batch = self.observations1.get_batch(batch_idxs)
+        action_batch = self.actions.get_batch(batch_idxs)
+        reward_batch = self.rewards.get_batch(batch_idxs)
+        terminal1_batch = self.terminals1.get_batch(batch_idxs)
+
+        result = {
+            'obs0': array_min2d(obs0_batch),
+            'obs1': array_min2d(obs1_batch),
+            'rewards': array_min2d(reward_batch),
+            'actions': array_min2d(action_batch),
+            'terminals1': array_min2d(terminal1_batch),
+        }
+        return result
+
+    def append(self, obs0, action, reward, obs1, terminal1, training=True):
+        if not training:
+            return
+        
+        self.observations0.append(obs0)
+        self.actions.append(action)
+        self.rewards.append(reward)
+        self.observations1.append(obs1)
+        self.terminals1.append(terminal1)
+
+    @property
+    def nb_entries(self):
+        return len(self.observations0)

baselines/ddpg/models.py

@@ -0,0 +1,77 @@
+import tensorflow as tf
+import tensorflow.contrib as tc
+
+
+class Model(object):
+    def __init__(self, name):
+        self.name = name
+
+    @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_vars if 'LayerNorm' not in var.name]
+
+
+class Actor(Model):
+    def __init__(self, nb_actions, name='actor', layer_norm=True):
+        super(Actor, self).__init__(name=name)
+        self.nb_actions = nb_actions
+        self.layer_norm = layer_norm
+
+    def __call__(self, obs, reuse=False):
+        with tf.variable_scope(self.name) as scope:
+            if reuse:
+                scope.reuse_variables()
+
+            x = obs
+            x = tf.layers.dense(x, 64)
+            if self.layer_norm:
+                x = tc.layers.layer_norm(x, center=True, scale=True)
+            x = tf.nn.relu(x)
+            
+            x = tf.layers.dense(x, 64)
+            if self.layer_norm:
+                x = tc.layers.layer_norm(x, center=True, scale=True)
+            x = tf.nn.relu(x)
+            
+            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, name='critic', layer_norm=True):
+        super(Critic, self).__init__(name=name)
+        self.layer_norm = layer_norm
+
+    def __call__(self, obs, action, reuse=False):
+        with tf.variable_scope(self.name) as scope:
+            if reuse:
+                scope.reuse_variables()
+
+            x = obs
+            x = tf.layers.dense(x, 64)
+            if self.layer_norm:
+                x = tc.layers.layer_norm(x, center=True, scale=True)
+            x = tf.nn.relu(x)
+
+            x = tf.concat([x, action], axis=-1)
+            x = tf.layers.dense(x, 64)
+            if self.layer_norm:
+                x = tc.layers.layer_norm(x, center=True, scale=True)
+            x = tf.nn.relu(x)
+
+            x = tf.layers.dense(x, 1, kernel_initializer=tf.random_uniform_initializer(minval=-3e-3, maxval=3e-3))
+        return x
+
+    @property
+    def output_vars(self):
+        output_vars = [var for var in self.trainable_vars if 'output' in var.name]
+        return output_vars

baselines/ddpg/noise.py

@@ -0,0 +1,67 @@
+import numpy as np
+
+
+class AdaptiveParamNoiseSpec(object):
+    def __init__(self, initial_stddev=0.1, desired_action_stddev=0.1, adoption_coefficient=1.01):
+        self.initial_stddev = initial_stddev
+        self.desired_action_stddev = desired_action_stddev
+        self.adoption_coefficient = adoption_coefficient
+
+        self.current_stddev = initial_stddev
+
+    def adapt(self, distance):
+        if distance > self.desired_action_stddev:
+            # Decrease stddev.
+            self.current_stddev /= self.adoption_coefficient
+        else:
+            # Increase stddev.
+            self.current_stddev *= self.adoption_coefficient
+
+    def get_stats(self):
+        stats = {
+            'param_noise_stddev': self.current_stddev,
+        }
+        return stats
+
+    def __repr__(self):
+        fmt = 'AdaptiveParamNoiseSpec(initial_stddev={}, desired_action_stddev={}, adoption_coefficient={})'
+        return fmt.format(self.initial_stddev, self.desired_action_stddev, self.adoption_coefficient)
+
+
+class ActionNoise(object):
+    def reset(self):
+        pass
+
+
+class NormalActionNoise(ActionNoise):
+    def __init__(self, mu, sigma):
+        self.mu = mu
+        self.sigma = sigma
+
+    def __call__(self):
+        return np.random.normal(self.mu, self.sigma)
+
+    def __repr__(self):
+        return 'NormalActionNoise(mu={}, sigma={})'.format(self.mu, self.sigma)
+
+
+# Based on http://math.stackexchange.com/questions/1287634/implementing-ornstein-uhlenbeck-in-matlab
+class OrnsteinUhlenbeckActionNoise(ActionNoise):
+    def __init__(self, mu, sigma, theta=.15, dt=1e-2, x0=None):
+        self.theta = theta
+        self.mu = mu
+        self.sigma = sigma
+        self.dt = dt
+        self.x0 = x0
+        self.reset()
+
+    def __call__(self):
+        x = self.x_prev + self.theta * (self.mu - self.x_prev) * self.dt + self.sigma * np.sqrt(self.dt) * np.random.normal(size=self.mu.shape)
+        self.x_prev = x
+        return x
+
+    def reset(self):
+        self.x_prev = self.x0 if self.x0 is not None else np.zeros_like(self.mu)
+
+    def __repr__(self):
+        return 'OrnsteinUhlenbeckActionNoise(mu={}, sigma={})'.format(self.mu, self.sigma)

baselines/ddpg/training.py

@@ -0,0 +1,189 @@
+import os
+import time
+from collections import deque
+import pickle
+
+from baselines.ddpg.ddpg import DDPG
+from baselines.ddpg.util import mpi_mean, mpi_std, mpi_max, mpi_sum
+import baselines.common.tf_util as U
+
+from baselines import logger
+import numpy as np
+import tensorflow as tf
+from mpi4py import MPI
+
+
+def train(env, nb_epochs, nb_epoch_cycles, render_eval, reward_scale, render, param_noise, actor, critic,
+    normalize_returns, normalize_observations, critic_l2_reg, actor_lr, critic_lr, action_noise,
+    popart, gamma, clip_norm, nb_train_steps, nb_rollout_steps, nb_eval_steps, batch_size, memory,
+    tau=0.01, eval_env=None, param_noise_adaption_interval=50):
+    rank = MPI.COMM_WORLD.Get_rank()
+
+    assert (np.abs(env.action_space.low) == env.action_space.high).all()  # we assume symmetric actions.
+    max_action = env.action_space.high
+    logger.info('scaling actions by {} before executing in env'.format(max_action))
+    agent = DDPG(actor, critic, memory, env.observation_space.shape, env.action_space.shape,
+        gamma=gamma, tau=tau, normalize_returns=normalize_returns, normalize_observations=normalize_observations,
+        batch_size=batch_size, action_noise=action_noise, param_noise=param_noise, critic_l2_reg=critic_l2_reg,
+        actor_lr=actor_lr, critic_lr=critic_lr, enable_popart=popart, clip_norm=clip_norm,
+        reward_scale=reward_scale)
+    logger.info('Using agent with the following configuration:')
+    logger.info(str(agent.__dict__.items()))
+
+    # Set up logging stuff only for a single worker.
+    if rank == 0:
+        saver = tf.train.Saver()
+    else:
+        saver = None
+    
+    step = 0
+    episode = 0
+    eval_episode_rewards_history = deque(maxlen=100)
+    episode_rewards_history = deque(maxlen=100)
+    with U.single_threaded_session() as sess:
+        # Prepare everything.
+        agent.initialize(sess)
+        sess.graph.finalize()
+
+        agent.reset()
+        obs = env.reset()
+        if eval_env is not None:
+            eval_obs = eval_env.reset()
+        done = False
+        episode_reward = 0.
+        episode_step = 0
+        episodes = 0
+        t = 0
+
+        epoch = 0
+        start_time = time.time()
+
+        epoch_episode_rewards = []
+        epoch_episode_steps = []
+        epoch_episode_eval_rewards = []
+        epoch_episode_eval_steps = []
+        epoch_start_time = time.time()
+        epoch_actions = []
+        epoch_qs = []
+        epoch_episodes = 0
+        for epoch in range(nb_epochs):
+            for cycle in range(nb_epoch_cycles):
+                # Perform rollouts.
+                for t_rollout in range(nb_rollout_steps):
+                    # Predict next action.
+                    action, q = agent.pi(obs, apply_noise=True, compute_Q=True)
+                    assert action.shape == env.action_space.shape
+
+                    # Execute next action.
+                    if rank == 0 and render:
+                        env.render()
+                    assert max_action.shape == action.shape
+                    new_obs, r, done, info = env.step(max_action * action)  # scale for execution in env (as far as DDPG is concerned, every action is in [-1, 1])
+                    t += 1
+                    if rank == 0 and render:
+                        env.render()
+                    episode_reward += r
+                    episode_step += 1
+
+                    # Book-keeping.
+                    epoch_actions.append(action)
+                    epoch_qs.append(q)
+                    agent.store_transition(obs, action, r, new_obs, done)
+                    obs = new_obs
+
+                    if done:
+                        # Episode done.
+                        epoch_episode_rewards.append(episode_reward)
+                        episode_rewards_history.append(episode_reward)
+                        epoch_episode_steps.append(episode_step)
+                        episode_reward = 0.
+                        episode_step = 0
+                        epoch_episodes += 1
+                        episodes += 1
+
+                        agent.reset()
+                        obs = env.reset()
+
+                # Train.
+                epoch_actor_losses = []
+                epoch_critic_losses = []
+                epoch_adaptive_distances = []
+                for t_train in range(nb_train_steps):
+                    # Adapt param noise, if necessary.
+                    if memory.nb_entries >= batch_size and t % param_noise_adaption_interval == 0:
+                        distance = agent.adapt_param_noise()
+                        epoch_adaptive_distances.append(distance)
+
+                    cl, al = agent.train()
+                    epoch_critic_losses.append(cl)
+                    epoch_actor_losses.append(al)
+                    agent.update_target_net()
+
+                # Evaluate.
+                eval_episode_rewards = []
+                eval_qs = []
+                if eval_env is not None:
+                    eval_episode_reward = 0.
+                    for t_rollout in range(nb_eval_steps):
+                        eval_action, eval_q = agent.pi(eval_obs, apply_noise=False, compute_Q=True)
+                        eval_obs, eval_r, eval_done, eval_info = eval_env.step(max_action * eval_action)  # scale for execution in env (as far as DDPG is concerned, every action is in [-1, 1])
+                        if render_eval:
+                            eval_env.render()
+                        eval_episode_reward += eval_r
+
+                        eval_qs.append(eval_q)
+                        if eval_done:
+                            eval_obs = eval_env.reset()
+                            eval_episode_rewards.append(eval_episode_reward)
+                            eval_episode_rewards_history.append(eval_episode_reward)
+                            eval_episode_reward = 0.
+
+            # Log stats.
+            epoch_train_duration = time.time() - epoch_start_time
+            duration = time.time() - start_time
+            stats = agent.get_stats()
+            combined_stats = {}
+            for key in sorted(stats.keys()):
+                combined_stats[key] = mpi_mean(stats[key])
+
+            # Rollout statistics.
+            combined_stats['rollout/return'] = mpi_mean(epoch_episode_rewards)
+            combined_stats['rollout/return_history'] = mpi_mean(np.mean(episode_rewards_history))
+            combined_stats['rollout/episode_steps'] = mpi_mean(epoch_episode_steps)
+            combined_stats['rollout/episodes'] = mpi_sum(epoch_episodes)
+            combined_stats['rollout/actions_mean'] = mpi_mean(epoch_actions)
+            combined_stats['rollout/actions_std'] = mpi_std(epoch_actions)
+            combined_stats['rollout/Q_mean'] = mpi_mean(epoch_qs)
+    
+            # Train statistics.
+            combined_stats['train/loss_actor'] = mpi_mean(epoch_actor_losses)
+            combined_stats['train/loss_critic'] = mpi_mean(epoch_critic_losses)
+            combined_stats['train/param_noise_distance'] = mpi_mean(epoch_adaptive_distances)
+
+            # Evaluation statistics.
+            if eval_env is not None:
+                combined_stats['eval/return'] = mpi_mean(eval_episode_rewards)
+                combined_stats['eval/return_history'] = mpi_mean(np.mean(eval_episode_rewards_history))
+                combined_stats['eval/Q'] = mpi_mean(eval_qs)
+                combined_stats['eval/episodes'] = mpi_mean(len(eval_episode_rewards))
+
+            # Total statistics.
+            combined_stats['total/duration'] = mpi_mean(duration)
+            combined_stats['total/steps_per_second'] = mpi_mean(float(t) / float(duration))
+            combined_stats['total/episodes'] = mpi_mean(episodes)
+            combined_stats['total/epochs'] = epoch + 1
+            combined_stats['total/steps'] = t
+            
+            for key in sorted(combined_stats.keys()):
+                logger.record_tabular(key, combined_stats[key])
+            logger.dump_tabular()
+            logger.info('')
+            logdir = logger.get_dir()
+            if rank == 0 and logdir:
+                if hasattr(env, 'get_state'):
+                    with open(os.path.join(logdir, 'env_state.pkl'), 'wb') as f:
+                        pickle.dump(env.get_state(), f)
+                if eval_env and hasattr(eval_env, 'get_state'):
+                    with open(os.path.join(logdir, 'eval_env_state.pkl'), 'wb') as f:
+                        pickle.dump(eval_env.get_state(), f)
+

baselines/ddpg/util.py

@@ -0,0 +1,44 @@
+import numpy as np
+import tensorflow as tf
+from mpi4py import MPI
+from baselines.common.mpi_moments import mpi_moments
+
+
+def reduce_var(x, axis=None, keepdims=False):
+    m = tf.reduce_mean(x, axis=axis, keep_dims=True)
+    devs_squared = tf.square(x - m)
+    return tf.reduce_mean(devs_squared, axis=axis, keep_dims=keepdims)
+
+
+def reduce_std(x, axis=None, keepdims=False):
+    return tf.sqrt(reduce_var(x, axis=axis, keepdims=keepdims))
+
+
+def mpi_mean(value):
+    if value == []:
+        value = [0.]
+    if not isinstance(value, list):
+        value = [value]
+    return mpi_moments(np.array(value))[0][0]
+
+
+def mpi_std(value):
+    if value == []:
+        value = [0.]
+    if not isinstance(value, list):
+        value = [value]
+    return mpi_moments(np.array(value))[1][0]
+
+
+def mpi_max(value):
+    global_max = np.zeros(1, dtype='float64')
+    local_max = np.max(value).astype('float64')
+    MPI.COMM_WORLD.Reduce(local_max, global_max, op=MPI.MAX)
+    return global_max[0]
+
+
+def mpi_sum(value):
+    global_sum = np.zeros(1, dtype='float64')
+    local_sum = np.sum(np.array(value)).astype('float64')
+    MPI.COMM_WORLD.Reduce(local_sum, global_sum, op=MPI.SUM)
+    return global_sum[0]

baselines/deepq/README.md

@@ -21,8 +21,8 @@ Be sure to check out the source code of [both](experiments/train_cartpole.py) [f
 
 Check out our simple agent trained with one stop shop `deepq.learn` function. 
 
-- `baselines/deepq/experiments/train_cartpole.py` - train a Cartpole agent.
-- `baselines/deepq/experiments/train_pong.py` - train a Pong agent using convolutional neural networks.
+- [baselines/deepq/experiments/train_cartpole.py](experiments/train_cartpole.py) - train a Cartpole agent.
+- [baselines/deepq/experiments/train_pong.py](experiments/train_pong.py) - train a Pong agent using convolutional neural networks.
 
 In particular notice that once `deepq.learn` finishes training it returns `act` function which can be used to select actions in the environment. Once trained you can easily save it and load at later time. For both of the files listed above there are complimentary files `enjoy_cartpole.py` and `enjoy_pong.py` respectively, that load and visualize the learned policy.
 
@@ -31,8 +31,8 @@ In particular notice that once `deepq.learn` finishes training it returns `act`
 ##### Check out the examples
 
 
-- `baselines/deepq/experiments/custom_cartpole.py` - Cartpole training with more fine grained control over the internals of DQN algorithm.
-- `baselines/deepq/experiments/atari/train.py` - more robust setup for training at scale.
+- [baselines/deepq/experiments/custom_cartpole.py](experiments/custom_cartpole.py) - Cartpole training with more fine grained control over the internals of DQN algorithm.
+- [baselines/deepq/experiments/atari/train.py](experiments/atari/train.py) - more robust setup for training at scale.
 
 
 ##### Download a pretrained Atari agent

baselines/deepq/__init__.py

@@ -1,5 +1,8 @@
 from baselines.deepq import models  # noqa
 from baselines.deepq.build_graph import build_act, build_train  # noqa
-
 from baselines.deepq.simple import learn, load  # noqa
 from baselines.deepq.replay_buffer import ReplayBuffer, PrioritizedReplayBuffer  # noqa
+
+def wrap_atari_dqn(env):
+    from baselines.common.atari_wrappers import wrap_deepmind
+    return wrap_deepmind(env, frame_stack=True, scale=True)

baselines/deepq/build_graph.py

@@ -22,6 +22,32 @@ The functions in this file can are used to create the following functions:
     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 a new value, if negative not 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:
@@ -71,6 +97,21 @@ import tensorflow as tf
 import baselines.common.tf_util as U
 
 
+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:
 
@@ -118,7 +159,6 @@ def build_act(make_obs_ph, q_func, num_actions, scope="deepq", reuse=None):
 
         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},
@@ -126,7 +166,121 @@ def build_act(make_obs_ph, q_func, num_actions, scope="deepq", reuse=None):
         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):
+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 = U.ensure_tf_input(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 = U.scope_vars(U.absolute_scope_name("q_func"))
+            all_perturbed_vars = U.scope_vars(U.absolute_scope_name("perturbed_q_func"))
+            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)
+        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
@@ -160,6 +314,11 @@ def build_train(make_obs_ph, q_func, num_actions, optimizer, grad_norm_clipping=
         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
     -------
@@ -175,7 +334,11 @@ def build_train(make_obs_ph, q_func, num_actions, optimizer, grad_norm_clipping=
     debug: {str: function}
         a bunch of functions to print debug data like q_values.
     """
-    act_f = build_act(make_obs_ph, q_func, num_actions, scope=scope, reuse=reuse)
+    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
@@ -213,6 +376,7 @@ def build_train(make_obs_ph, q_func, num_actions, optimizer, grad_norm_clipping=
         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:
             optimize_expr = U.minimize_and_clip(optimizer,

baselines/deepq/experiments/atari/enjoy.py

@@ -10,8 +10,8 @@ import baselines.common.tf_util as U
 from baselines import deepq
 from baselines.common.misc_util import (
     boolean_flag,
-    SimpleMonitor,
 )
+from baselines import bench
 from baselines.common.atari_wrappers_deprecated import wrap_dqn
 from baselines.deepq.experiments.atari.model import model, dueling_model
 
@@ -30,7 +30,7 @@ def parse_args():
 
 def make_env(game_name):
     env = gym.make(game_name + "NoFrameskip-v4")
-    env = SimpleMonitor(env)
+    env = bench.Monitor(env, None)
     env = wrap_dqn(env)
     return env
 

baselines/deepq/experiments/atari/model.py

@@ -2,7 +2,14 @@ import tensorflow as tf
 import tensorflow.contrib.layers as layers
 
 
-def model(img_in, num_actions, scope, reuse=False):
+def layer_norm_fn(x, relu=True):
+    x = layers.layer_norm(x, scale=True, center=True)
+    if relu:
+        x = tf.nn.relu(x)
+    return x
+
+
+def model(img_in, num_actions, scope, reuse=False, layer_norm=False):
     """As described in https://storage.googleapis.com/deepmind-data/assets/papers/DeepMindNature14236Paper.pdf"""
     with tf.variable_scope(scope, reuse=reuse):
         out = img_in
@@ -11,16 +18,19 @@ def model(img_in, num_actions, scope, reuse=False):
             out = layers.convolution2d(out, num_outputs=32, kernel_size=8, stride=4, activation_fn=tf.nn.relu)
             out = layers.convolution2d(out, num_outputs=64, kernel_size=4, stride=2, activation_fn=tf.nn.relu)
             out = layers.convolution2d(out, num_outputs=64, kernel_size=3, stride=1, activation_fn=tf.nn.relu)
-        out = layers.flatten(out)
+        conv_out = layers.flatten(out)
 
         with tf.variable_scope("action_value"):
-            out = layers.fully_connected(out, num_outputs=512, activation_fn=tf.nn.relu)
-            out = layers.fully_connected(out, num_outputs=num_actions, activation_fn=None)
-
-        return out
+            value_out = layers.fully_connected(conv_out, num_outputs=512, activation_fn=None)
+            if layer_norm:
+                value_out = layer_norm_fn(value_out, relu=True)
+            else:
+                value_out = tf.nn.relu(value_out)
+            value_out = layers.fully_connected(value_out, num_outputs=num_actions, activation_fn=None)
+        return value_out
 
 
-def dueling_model(img_in, num_actions, scope, reuse=False):
+def dueling_model(img_in, num_actions, scope, reuse=False, layer_norm=False):
     """As described in https://arxiv.org/abs/1511.06581"""
     with tf.variable_scope(scope, reuse=reuse):
         out = img_in
@@ -29,15 +39,22 @@ def dueling_model(img_in, num_actions, scope, reuse=False):
             out = layers.convolution2d(out, num_outputs=32, kernel_size=8, stride=4, activation_fn=tf.nn.relu)
             out = layers.convolution2d(out, num_outputs=64, kernel_size=4, stride=2, activation_fn=tf.nn.relu)
             out = layers.convolution2d(out, num_outputs=64, kernel_size=3, stride=1, activation_fn=tf.nn.relu)
-        out = layers.flatten(out)
+        conv_out = layers.flatten(out)
 
         with tf.variable_scope("state_value"):
-            state_hidden = layers.fully_connected(out, num_outputs=512, activation_fn=tf.nn.relu)
+            state_hidden = layers.fully_connected(conv_out, num_outputs=512, activation_fn=None)
+            if layer_norm:
+                state_hidden = layer_norm_fn(state_hidden, relu=True)
+            else:
+                state_hidden = tf.nn.relu(state_hidden)
             state_score = layers.fully_connected(state_hidden, num_outputs=1, activation_fn=None)
         with tf.variable_scope("action_value"):
-            actions_hidden = layers.fully_connected(out, num_outputs=512, activation_fn=tf.nn.relu)
+            actions_hidden = layers.fully_connected(conv_out, num_outputs=512, activation_fn=None)
+            if layer_norm:
+                actions_hidden = layer_norm_fn(actions_hidden, relu=True)
+            else:
+                actions_hidden = tf.nn.relu(actions_hidden)
             action_scores = layers.fully_connected(actions_hidden, num_outputs=num_actions, activation_fn=None)
             action_scores_mean = tf.reduce_mean(action_scores, 1)
             action_scores = action_scores - tf.expand_dims(action_scores_mean, 1)
-
         return state_score + action_scores

baselines/deepq/experiments/atari/train.py

@@ -5,6 +5,7 @@ import os
 import tensorflow as tf
 import tempfile
 import time
+import json
 
 import baselines.common.tf_util as U
 
@@ -18,11 +19,9 @@ from baselines.common.misc_util import (
     relatively_safe_pickle_dump,
     set_global_seeds,
     RunningAvg,
-    SimpleMonitor
 )
 from baselines.common.schedules import LinearSchedule, PiecewiseSchedule
-# when updating this to non-deperecated ones, it is important to
-# copy over LazyFrames
+from baselines import bench
 from baselines.common.atari_wrappers_deprecated import wrap_dqn
 from baselines.common.azure_utils import Container
 from .model import model, dueling_model
@@ -40,10 +39,15 @@ def parse_args():
     parser.add_argument("--batch-size", type=int, default=32, help="number of transitions to optimize at the same time")
     parser.add_argument("--learning-freq", type=int, default=4, help="number of iterations between every optimization step")
     parser.add_argument("--target-update-freq", type=int, default=40000, help="number of iterations between every target network update")
+    parser.add_argument("--param-noise-update-freq", type=int, default=50, help="number of iterations between every re-scaling of the parameter noise")
+    parser.add_argument("--param-noise-reset-freq", type=int, default=10000, help="maximum number of steps to take per episode before re-perturbing the exploration policy")
     # Bells and whistles
     boolean_flag(parser, "double-q", default=True, help="whether or not to use double q learning")
     boolean_flag(parser, "dueling", default=False, help="whether or not to use dueling model")
     boolean_flag(parser, "prioritized", default=False, help="whether or not to use prioritized replay buffer")
+    boolean_flag(parser, "param-noise", default=False, help="whether or not to use parameter space noise for exploration")
+    boolean_flag(parser, "layer-norm", default=False, help="whether or not to use layer norm (should be True if param_noise is used)")
+    boolean_flag(parser, "gym-monitor", default=False, help="whether or not to use a OpenAI Gym monitor (results in slower training due to video recording)")
     parser.add_argument("--prioritized-alpha", type=float, default=0.6, help="alpha parameter for prioritized replay buffer")
     parser.add_argument("--prioritized-beta0", type=float, default=0.4, help="initial value of beta parameters for prioritized replay")
     parser.add_argument("--prioritized-eps", type=float, default=1e-6, help="eps parameter for prioritized replay buffer")
@@ -58,7 +62,7 @@ def parse_args():
 
 def make_env(game_name):
     env = gym.make(game_name + "NoFrameskip-v4")
-    monitored_env = SimpleMonitor(env)  # puts rewards and number of steps in info, before environment is wrapped
+    monitored_env = bench.Monitor(env, logger.get_dir())  # puts rewards and number of steps in info, before environment is wrapped
     env = wrap_dqn(monitored_env)  # applies a bunch of modification to simplify the observation space (downsample, make b/w)
     return env, monitored_env
 
@@ -104,8 +108,11 @@ def maybe_load_model(savedir, container):
 
 if __name__ == '__main__':
     args = parse_args()
+    
     # Parse savedir and azure container.
     savedir = args.save_dir
+    if savedir is None:
+        savedir = os.getenv('OPENAI_LOGDIR', None)
     if args.save_azure_container is not None:
         account_name, account_key, container_name = args.save_azure_container.split(":")
         container = Container(account_name=account_name,
@@ -123,16 +130,27 @@ if __name__ == '__main__':
         set_global_seeds(args.seed)
         env.unwrapped.seed(args.seed)
 
+    if args.gym_monitor and savedir:
+        env = gym.wrappers.Monitor(env, os.path.join(savedir, 'gym_monitor'), force=True)
+
+    if savedir:
+        with open(os.path.join(savedir, 'args.json'), 'w') as f:
+            json.dump(vars(args), f)
+
     with U.make_session(4) as sess:
         # Create training graph and replay buffer
+        def model_wrapper(img_in, num_actions, scope, **kwargs):
+            actual_model = dueling_model if args.dueling else model
+            return actual_model(img_in, num_actions, scope, layer_norm=args.layer_norm, **kwargs)
         act, train, update_target, debug = deepq.build_train(
             make_obs_ph=lambda name: U.Uint8Input(env.observation_space.shape, name=name),
-            q_func=dueling_model if args.dueling else model,
+            q_func=model_wrapper,
             num_actions=env.action_space.n,
             optimizer=tf.train.AdamOptimizer(learning_rate=args.lr, epsilon=1e-4),
             gamma=0.99,
             grad_norm_clipping=10,
-            double_q=args.double_q
+            double_q=args.double_q,
+            param_noise=args.param_noise
         )
 
         approximate_num_iters = args.num_steps / 4
@@ -162,17 +180,43 @@ if __name__ == '__main__':
         steps_per_iter = RunningAvg(0.999)
         iteration_time_est = RunningAvg(0.999)
         obs = env.reset()
+        num_iters_since_reset = 0
+        reset = True
 
         # Main trianing loop
         while True:
             num_iters += 1
+            num_iters_since_reset += 1
+
             # Take action and store transition in the replay buffer.
-            action = act(np.array(obs)[None], update_eps=exploration.value(num_iters))[0]
+            kwargs = {}
+            if not args.param_noise:
+                update_eps = exploration.value(num_iters)
+                update_param_noise_threshold = 0.
+            else:
+                if args.param_noise_reset_freq > 0 and num_iters_since_reset > args.param_noise_reset_freq:
+                    # Reset param noise policy since we have exceeded the maximum number of steps without a reset.
+                    reset = True
+
+                update_eps = 0.01  # ensures that we cannot get stuck completely
+                # 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(num_iters) + exploration.value(num_iters) / float(env.action_space.n))
+                kwargs['reset'] = reset
+                kwargs['update_param_noise_threshold'] = update_param_noise_threshold
+                kwargs['update_param_noise_scale'] = (num_iters % args.param_noise_update_freq == 0)
+
+            action = act(np.array(obs)[None], update_eps=update_eps, **kwargs)[0]
+            reset = False
             new_obs, rew, done, info = env.step(action)
             replay_buffer.add(obs, action, rew, new_obs, float(done))
             obs = new_obs
             if done:
+                num_iters_since_reset = 0
                 obs = env.reset()
+                reset = True
 
             if (num_iters > max(5 * args.batch_size, args.replay_buffer_size // 20) and
                     num_iters % args.learning_freq == 0):
@@ -203,7 +247,7 @@ if __name__ == '__main__':
                 maybe_save_model(savedir, container, {
                     'replay_buffer': replay_buffer,
                     'num_iters': num_iters,
-                    'monitor_state': monitored_env.get_state()
+                    'monitor_state': monitored_env.get_state(),
                 })
 
             if info["steps"] > args.num_steps:

baselines/deepq/experiments/atari/wang2015_eval.py

@@ -5,15 +5,15 @@ import os
 
 import baselines.common.tf_util as U
 
-from baselines import deepq
-from baselines.common.misc_util import get_wrapper_by_name, SimpleMonitor, boolean_flag, set_global_seeds
+from baselines import deepq, bench
+from baselines.common.misc_util import get_wrapper_by_name, boolean_flag, set_global_seeds
 from baselines.common.atari_wrappers_deprecated import wrap_dqn
 from baselines.deepq.experiments.atari.model import model, dueling_model
 
 
 def make_env(game_name):
     env = gym.make(game_name + "NoFrameskip-v4")
-    env_monitored = SimpleMonitor(env)
+    env_monitored = bench.Monitor(env, None)
     env = wrap_dqn(env_monitored)
     return env_monitored, env
 
@@ -47,14 +47,14 @@ def wang2015_eval(game_name, act, stochastic):
             eval_episode_steps += 1
             action = act(np.array(obs)[None], stochastic=stochastic)[0]
 
-            obs, reward, done, info = eval_env.step(action)
+            obs, _reward, done, info = eval_env.step(action)
             if done:
                 obs = eval_env.reset()
             if len(info["rewards"]) > 0:
                 episode_rewards.append(info["rewards"][0])
                 break
             if info["steps"] > 108000:  # 5 minutes of gameplay
-                episode_rewards.append(env_monitored._current_reward)
+                episode_rewards.append(sum(env_monitored.rewards))
                 break
         print("Num steps in episode {} was {} yielding {} reward".format(
               num_noops, eval_episode_steps, episode_rewards[-1]), flush=True)
@@ -66,7 +66,7 @@ def wang2015_eval(game_name, act, stochastic):
 def main():
     set_global_seeds(1)
     args = parse_args()
-    with U.make_session(4) as sess:  # noqa
+    with U.make_session(4):  # noqa
         _, env = make_env(args.env)
         act = deepq.build_act(
             make_obs_ph=lambda name: U.Uint8Input(env.observation_space.shape, name=name),

baselines/deepq/experiments/enjoy_mountaincar.py

@@ -0,0 +1,21 @@
+import gym
+
+from baselines import deepq
+
+
+def main():
+    env = gym.make("MountainCar-v0")
+    act = deepq.load("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

@@ -1,12 +1,10 @@
 import gym
-
 from baselines import deepq
-from baselines.common.atari_wrappers_deprecated import wrap_dqn, ScaledFloatFrame
 
 
 def main():
     env = gym.make("PongNoFrameskip-v4")
-    env = ScaledFloatFrame(wrap_dqn(env))
+    env = deepq.wrap_atari_dqn(env)
     act = deepq.load("pong_model.pkl")
 
     while True:

baselines/deepq/experiments/run_atari.py

@@ -0,0 +1,47 @@
+import gym
+
+from baselines import deepq
+from baselines.common import set_global_seeds
+from baselines import bench
+import argparse
+from baselines import logger
+from baselines.common.atari_wrappers import make_atari
+
+def main():
+    parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter)
+    parser.add_argument('--env', help='environment ID', default='BreakoutNoFrameskip-v4')
+    parser.add_argument('--seed', help='RNG seed', type=int, default=0)
+    parser.add_argument('--prioritized', type=int, default=1)
+    parser.add_argument('--dueling', type=int, default=1)
+    parser.add_argument('--num-timesteps', type=int, default=int(10e6))
+    args = parser.parse_args()
+    logger.configure()
+    set_global_seeds(args.seed)
+    env = make_atari(args.env)
+    env = bench.Monitor(env, logger.get_dir())
+    env = deepq.wrap_atari_dqn(env)
+    model = deepq.models.cnn_to_mlp(
+        convs=[(32, 8, 4), (64, 4, 2), (64, 3, 1)],
+        hiddens=[256],
+        dueling=bool(args.dueling),
+    )
+    act = deepq.learn(
+        env,
+        q_func=model,
+        lr=1e-4,
+        max_timesteps=args.num_timesteps,
+        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,
+        prioritized_replay=bool(args.prioritized)
+    )
+    # act.save("pong_model.pkl") XXX
+    env.close()
+
+
+if __name__ == '__main__':
+    main()

baselines/deepq/experiments/train_mountaincar.py

@@ -0,0 +1,26 @@
+import gym
+
+from baselines import deepq
+
+
+def main():
+    env = gym.make("MountainCar-v0")
+    # Enabling layer_norm here is import for parameter space noise!
+    model = deepq.models.mlp([64], layer_norm=True)
+    act = deepq.learn(
+        env,
+        q_func=model,
+        lr=1e-3,
+        max_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

@@ -1,34 +0,0 @@
-import gym
-
-from baselines import deepq
-from baselines.common.atari_wrappers_deprecated import wrap_dqn, ScaledFloatFrame
-
-
-def main():
-    env = gym.make("PongNoFrameskip-v4")
-    env = ScaledFloatFrame(wrap_dqn(env))
-    model = deepq.models.cnn_to_mlp(
-        convs=[(32, 8, 4), (64, 4, 2), (64, 3, 1)],
-        hiddens=[256],
-        dueling=True
-    )
-    act = deepq.learn(
-        env,
-        q_func=model,
-        lr=1e-4,
-        max_timesteps=2000000,
-        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,
-        prioritized_replay=True
-    )
-    act.save("pong_model.pkl")
-    env.close()
-
-
-if __name__ == '__main__':
-    main()

baselines/deepq/models.py

@@ -2,16 +2,19 @@ import tensorflow as tf
 import tensorflow.contrib.layers as layers
 
 
-def _mlp(hiddens, inpt, num_actions, scope, reuse=False):
+def _mlp(hiddens, inpt, num_actions, scope, reuse=False, layer_norm=False):
     with tf.variable_scope(scope, reuse=reuse):
         out = inpt
         for hidden in hiddens:
-            out = layers.fully_connected(out, num_outputs=hidden, activation_fn=tf.nn.relu)
-        out = layers.fully_connected(out, num_outputs=num_actions, activation_fn=None)
-        return out
+            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=[]):
+def mlp(hiddens=[], layer_norm=False):
     """This model takes as input an observation and returns values of all actions.
 
     Parameters
@@ -24,10 +27,10 @@ def mlp(hiddens=[]):
     q_func: function
         q_function for DQN algorithm.
     """
-    return lambda *args, **kwargs: _mlp(hiddens, *args, **kwargs)
+    return lambda *args, **kwargs: _mlp(hiddens, layer_norm=layer_norm, *args, **kwargs)
 
 
-def _cnn_to_mlp(convs, hiddens, dueling, inpt, num_actions, scope, reuse=False):
+def _cnn_to_mlp(convs, hiddens, dueling, inpt, num_actions, scope, reuse=False, layer_norm=False):
     with tf.variable_scope(scope, reuse=reuse):
         out = inpt
         with tf.variable_scope("convnet"):
@@ -37,28 +40,34 @@ def _cnn_to_mlp(convs, hiddens, dueling, inpt, num_actions, scope, reuse=False):
                                            kernel_size=kernel_size,
                                            stride=stride,
                                            activation_fn=tf.nn.relu)
-        out = layers.flatten(out)
+        conv_out = layers.flatten(out)
         with tf.variable_scope("action_value"):
-            action_out = out
+            action_out = conv_out
             for hidden in hiddens:
-                action_out = layers.fully_connected(action_out, num_outputs=hidden, activation_fn=tf.nn.relu)
+                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 = out
+                state_out = conv_out
                 for hidden in hiddens:
-                    state_out = layers.fully_connected(state_out, num_outputs=hidden, activation_fn=tf.nn.relu)
+                    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)
-            return state_score + action_scores_centered
+            q_out = state_score + action_scores_centered
         else:
-            return action_scores
-        return out
+            q_out = action_scores
+        return q_out
 
 
-def cnn_to_mlp(convs, hiddens, dueling=False):
+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
@@ -78,5 +87,5 @@ def cnn_to_mlp(convs, hiddens, dueling=False):
         q_function for DQN algorithm.
     """
 
-    return lambda *args, **kwargs: _cnn_to_mlp(convs, hiddens, dueling, *args, **kwargs)
+    return lambda *args, **kwargs: _cnn_to_mlp(convs, hiddens, dueling, layer_norm=layer_norm, *args, **kwargs)
 

baselines/deepq/simple.py

@@ -1,12 +1,13 @@
-import numpy as np
 import os
-import dill
 import tempfile
+
 import tensorflow as tf
 import zipfile
+import cloudpickle
+import numpy as np
 
+import gym
 import baselines.common.tf_util as U
-
 from baselines import logger
 from baselines.common.schedules import LinearSchedule
 from baselines import deepq
@@ -19,11 +20,11 @@ class ActWrapper(object):
         self._act_params = act_params
 
     @staticmethod
-    def load(path, num_cpu=16):
+    def load(path):
         with open(path, "rb") as f:
-            model_data, act_params = dill.load(f)
+            model_data, act_params = cloudpickle.load(f)
         act = deepq.build_act(**act_params)
-        sess = U.make_session(num_cpu=num_cpu)
+        sess = tf.Session()
         sess.__enter__()
         with tempfile.TemporaryDirectory() as td:
             arc_path = os.path.join(td, "packed.zip")
@@ -38,8 +39,11 @@ class ActWrapper(object):
     def __call__(self, *args, **kwargs):
         return self._act(*args, **kwargs)
 
-    def save(self, path):
+    def save(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:
             U.save_state(os.path.join(td, "model"))
             arc_name = os.path.join(td, "packed.zip")
@@ -52,18 +56,16 @@ class ActWrapper(object):
             with open(arc_name, "rb") as f:
                 model_data = f.read()
         with open(path, "wb") as f:
-            dill.dump((model_data, self._act_params), f)
+            cloudpickle.dump((model_data, self._act_params), f)
 
 
-def load(path, num_cpu=16):
+def load(path):
     """Load act function that was returned by learn function.
 
     Parameters
     ----------
     path: str
         path to the act function pickle
-    num_cpu: int
-        number of cpus to use for executing the policy
 
     Returns
     -------
@@ -71,7 +73,7 @@ def load(path, num_cpu=16):
         function that takes a batch of observations
         and returns actions.
     """
-    return ActWrapper.load(path, num_cpu=num_cpu)
+    return ActWrapper.load(path)
 
 
 def learn(env,
@@ -83,7 +85,7 @@ def learn(env,
           exploration_final_eps=0.02,
           train_freq=1,
           batch_size=32,
-          print_freq=1,
+          print_freq=100,
           checkpoint_freq=10000,
           learning_starts=1000,
           gamma=1.0,
@@ -93,7 +95,7 @@ def learn(env,
           prioritized_replay_beta0=0.4,
           prioritized_replay_beta_iters=None,
           prioritized_replay_eps=1e-6,
-          num_cpu=16,
+          param_noise=False,
           callback=None):
     """Train a deepq model.
 
@@ -150,8 +152,6 @@ def learn(env,
         to 1.0. If set to None equals to max_timesteps.
     prioritized_replay_eps: float
         epsilon to add to the TD errors when updating priorities.
-    num_cpu: int
-        number of cpus to use for training
     callback: (locals, globals) -> None
         function called at every steps with state of the algorithm.
         If callback returns true training stops.
@@ -164,11 +164,14 @@ def learn(env,
     """
     # Create all the functions necessary to train the model
 
-    sess = U.make_session(num_cpu=num_cpu)
+    sess = tf.Session()
     sess.__enter__()
 
+    # capture the shape outside the closure so that the env object is not serialized
+    # by cloudpickle when serializing make_obs_ph
+    observation_space_shape = env.observation_space.shape
     def make_obs_ph(name):
-        return U.BatchInput(env.observation_space.shape, name=name)
+        return U.BatchInput(observation_space_shape, name=name)
 
     act, train, update_target, debug = deepq.build_train(
         make_obs_ph=make_obs_ph,
@@ -176,13 +179,18 @@ def learn(env,
         num_actions=env.action_space.n,
         optimizer=tf.train.AdamOptimizer(learning_rate=lr),
         gamma=gamma,
-        grad_norm_clipping=10
+        grad_norm_clipping=10,
+        param_noise=param_noise
     )
+
     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:
         replay_buffer = PrioritizedReplayBuffer(buffer_size, alpha=prioritized_replay_alpha)
@@ -206,6 +214,7 @@ def learn(env,
     episode_rewards = [0.0]
     saved_mean_reward = None
     obs = env.reset()
+    reset = True
     with tempfile.TemporaryDirectory() as td:
         model_saved = False
         model_file = os.path.join(td, "model")
@@ -214,8 +223,28 @@ def learn(env,
                 if callback(locals(), globals()):
                     break
             # Take action and update exploration to the newest value
-            action = act(np.array(obs)[None], update_eps=exploration.value(t))[0]
-            new_obs, rew, done, _ = env.step(action)
+            kwargs = {}
+            if not param_noise:
+                update_eps = exploration.value(t)
+                update_param_noise_threshold = 0.
+            else:
+                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]
+            if isinstance(env.action_space, gym.spaces.MultiBinary):
+                env_action = np.zeros(env.action_space.n)
+                env_action[action] = 1
+            else:
+                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
@@ -224,6 +253,7 @@ def learn(env,
             if done:
                 obs = env.reset()
                 episode_rewards.append(0.0)
+                reset = True
 
             if t > learning_starts and t % train_freq == 0:
                 # Minimize the error in Bellman's equation on a batch sampled from replay buffer.
@@ -265,4 +295,4 @@ def learn(env,
                 logger.log("Restored model with mean reward: {}".format(saved_mean_reward))
             U.load_state(model_file)
 
-    return ActWrapper(act, act_params)
+    return act

baselines/logger.py

@@ -1,13 +1,3 @@
-"""
-
-See README.md for a description of the logging API.
-
-OFF state corresponds to having Logger.CURRENT == Logger.DEFAULT
-ON state is otherwise
-
-"""
-
-from collections import OrderedDict
 import os
 import sys
 import shutil
@@ -16,8 +6,10 @@ import json
 import time
 import datetime
 import tempfile
+from mpi4py import MPI
 
-LOG_OUTPUT_FORMATS = ['stdout', 'log', 'json']
+LOG_OUTPUT_FORMATS = ['stdout', 'log', 'csv']
+# Also valid: json, tensorboard
 
 DEBUG = 10
 INFO = 20
@@ -26,42 +18,46 @@ ERROR = 40
 
 DISABLED = 50
 
-class OutputFormat(object):
+class KVWriter(object):
     def writekvs(self, kvs):
-        """
-        Write key-value pairs
-        """
         raise NotImplementedError
 
-    def writeseq(self, args):
-        """
-        Write a sequence of other data (e.g. a logging message)
-        """
-        pass
+class SeqWriter(object):
+    def writeseq(self, seq):
+        raise NotImplementedError
 
-    def close(self):
-        return
-
-
-class HumanOutputFormat(OutputFormat):
-    def __init__(self, file):
-        self.file = file
+class HumanOutputFormat(KVWriter, SeqWriter):
+    def __init__(self, filename_or_file):
+        if isinstance(filename_or_file, str):
+            self.file = open(filename_or_file, 'at')
+            self.own_file = True
+        else:
+            assert hasattr(filename_or_file, 'read'), 'expected file or str, got %s'%filename_or_file
+            self.file = filename_or_file
+            self.own_file = False
 
     def writekvs(self, kvs):
         # Create strings for printing
-        key2str = OrderedDict()
-        for (key, val) in kvs.items():
-            valstr = '%-8.3g' % (val,) if hasattr(val, '__float__') else val
+        key2str = {}
+        for (key, val) in sorted(kvs.items()):
+            if isinstance(val, float):
+                valstr = '%-8.3g' % (val,)
+            else:
+                valstr = str(val)
             key2str[self._truncate(key)] = self._truncate(valstr)
 
         # Find max widths
-        keywidth = max(map(len, key2str.keys()))
-        valwidth = max(map(len, key2str.values()))
+        if len(key2str) == 0:
+            print('WARNING: tried to write empty key-value dict')
+            return
+        else:
+            keywidth = max(map(len, key2str.keys()))
+            valwidth = max(map(len, key2str.values()))
 
         # Write out the data
         dashes = '-' * (keywidth + valwidth + 7)
         lines = [dashes]
-        for (key, val) in key2str.items():
+        for (key, val) in sorted(key2str.items()):
             lines.append('| %s%s | %s%s |' % (
                 key,
                 ' ' * (keywidth - len(key)),
@@ -77,25 +73,68 @@ class HumanOutputFormat(OutputFormat):
     def _truncate(self, s):
         return s[:20] + '...' if len(s) > 23 else s
 
-    def writeseq(self, args):
-        for arg in args:
+    def writeseq(self, seq):
+        for arg in seq:
             self.file.write(arg)
         self.file.write('\n')
         self.file.flush()
 
-class JSONOutputFormat(OutputFormat):
-    def __init__(self, file):
-        self.file = file
+    def close(self):
+        if self.own_file:
+            self.file.close()
+
+class JSONOutputFormat(KVWriter):
+    def __init__(self, filename):
+        self.file = open(filename, 'at')
 
     def writekvs(self, kvs):
-        for k, v in kvs.items():
+        for k, v in sorted(kvs.items()):
             if hasattr(v, 'dtype'):
                 v = v.tolist()
                 kvs[k] = float(v)
         self.file.write(json.dumps(kvs) + '\n')
         self.file.flush()
 
-class TensorBoardOutputFormat(OutputFormat):
+    def close(self):
+        self.file.close()
+
+class CSVOutputFormat(KVWriter):
+    def __init__(self, filename):
+        self.file = open(filename, 'a+t')
+        self.keys = []
+        self.sep = ','
+
+    def writekvs(self, kvs):
+        # Add our current row to the history
+        extra_keys = kvs.keys() - self.keys
+        if extra_keys:
+            self.keys.extend(extra_keys)
+            self.file.seek(0)
+            lines = self.file.readlines()
+            self.file.seek(0)
+            for (i, k) in enumerate(self.keys):
+                if i > 0:
+                    self.file.write(',')
+                self.file.write(k)
+            self.file.write('\n')
+            for line in lines[1:]:
+                self.file.write(line[:-1])
+                self.file.write(self.sep * len(extra_keys))
+                self.file.write('\n')
+        for (i, k) in enumerate(self.keys):
+            if i > 0:
+                self.file.write(',')
+            v = kvs.get(k)
+            if v:
+                self.file.write(str(v))
+        self.file.write('\n')
+        self.file.flush()
+
+    def close(self):
+        self.file.close()
+
+
+class TensorBoardOutputFormat(KVWriter):
     """
     Dumps key/value pairs into TensorBoard's numeric format.
     """
@@ -106,7 +145,7 @@ class TensorBoardOutputFormat(OutputFormat):
         prefix = 'events'
         path = osp.join(osp.abspath(dir), prefix)
         import tensorflow as tf
-        from tensorflow.python import pywrap_tensorflow        
+        from tensorflow.python import pywrap_tensorflow
         from tensorflow.core.util import event_pb2
         from tensorflow.python.util import compat
         self.tf = tf
@@ -130,18 +169,22 @@ class TensorBoardOutputFormat(OutputFormat):
             self.writer.Close()
             self.writer = None
 
-
 def make_output_format(format, ev_dir):
     os.makedirs(ev_dir, exist_ok=True)
+    rank = MPI.COMM_WORLD.Get_rank()
     if format == 'stdout':
         return HumanOutputFormat(sys.stdout)
     elif format == 'log':
-        log_file = open(osp.join(ev_dir, 'log.txt'), 'wt')
-        return HumanOutputFormat(log_file)
+        suffix = "" if rank==0 else ("-mpi%03i"%rank)
+        return HumanOutputFormat(osp.join(ev_dir, 'log%s.txt' % suffix))
     elif format == 'json':
-        json_file = open(osp.join(ev_dir, 'progress.json'), 'wt')
-        return JSONOutputFormat(json_file)
+        assert rank==0
+        return JSONOutputFormat(osp.join(ev_dir, 'progress.json'))
+    elif format == 'csv':
+        assert rank==0
+        return CSVOutputFormat(osp.join(ev_dir, 'progress.csv'))
     elif format == 'tensorboard':
+        assert rank==0
         return TensorBoardOutputFormat(osp.join(ev_dir, 'tb'))
     else:
         raise ValueError('Unknown format specified: %s' % (format,))
@@ -150,7 +193,6 @@ def make_output_format(format, ev_dir):
 # API
 # ================================================================
 
-
 def logkv(key, val):
     """
     Log a value of some diagnostic
@@ -158,6 +200,12 @@ def logkv(key, val):
     """
     Logger.CURRENT.logkv(key, val)
 
+def logkvs(d):
+    """
+    Log a dictionary of key-value pairs
+    """
+    for (k, v) in d.items():
+        logkv(k, v)
 
 def dumpkvs():
     """
@@ -168,10 +216,8 @@ def dumpkvs():
     """
     Logger.CURRENT.dumpkvs()
 
-
-# for backwards compatibility
-record_tabular = logkv
-dump_tabular = dumpkvs
+def getkvs():
+    return Logger.CURRENT.name2val
 
 
 def log(*args, level=INFO):
@@ -180,19 +226,15 @@ def log(*args, level=INFO):
     """
     Logger.CURRENT.log(*args, level=level)
 
-
 def debug(*args):
     log(*args, level=DEBUG)
 
-
 def info(*args):
     log(*args, level=INFO)
 
-
 def warn(*args):
     log(*args, level=WARN)
 
-
 def error(*args):
     log(*args, level=ERROR)
 
@@ -203,7 +245,6 @@ def set_level(level):
     """
     Logger.CURRENT.set_level(level)
 
-
 def get_dir():
     """
     Get directory that log files are being written to.
@@ -211,18 +252,20 @@ def get_dir():
     """
     return Logger.CURRENT.get_dir()
 
+record_tabular = logkv
+dump_tabular = dumpkvs
+
 # ================================================================
 # Backend
 # ================================================================
 
-
 class Logger(object):
     DEFAULT = None  # A logger with no output files. (See right below class definition)
                     # So that you can still log to the terminal without setting up any output files
     CURRENT = None  # Current logger being used by the free functions above
 
     def __init__(self, dir, output_formats):
-        self.name2val = OrderedDict()  # values this iteration
+        self.name2val = {}  # values this iteration
         self.level = INFO
         self.dir = dir
         self.output_formats = output_formats
@@ -233,8 +276,10 @@ class Logger(object):
         self.name2val[key] = val
 
     def dumpkvs(self):
+        if self.level == DISABLED: return
         for fmt in self.output_formats:
-            fmt.writekvs(self.name2val)
+            if isinstance(fmt, KVWriter):
+                fmt.writekvs(self.name2val)
         self.name2val.clear()
 
     def log(self, *args, level=INFO):
@@ -257,59 +302,48 @@ class Logger(object):
     # ----------------------------------------
     def _do_log(self, args):
         for fmt in self.output_formats:
-            fmt.writeseq(args)
+            if isinstance(fmt, SeqWriter):
+                fmt.writeseq(map(str, args))
 
+Logger.DEFAULT = Logger.CURRENT = Logger(dir=None, output_formats=[HumanOutputFormat(sys.stdout)])
 
-# ================================================================
+def configure(dir=None, format_strs=None):
+    if dir is None:
+        dir = os.getenv('OPENAI_LOGDIR')
+    if dir is None:
+        dir = osp.join(tempfile.gettempdir(),
+            datetime.datetime.now().strftime("openai-%Y-%m-%d-%H-%M-%S-%f"))
+    assert isinstance(dir, str)
+    os.makedirs(dir, exist_ok=True)
 
-Logger.DEFAULT = Logger(output_formats=[HumanOutputFormat(sys.stdout)], dir=None)
-Logger.CURRENT = Logger.DEFAULT
+    if format_strs is None:
+        strs = os.getenv('OPENAI_LOG_FORMAT')
+        format_strs = strs.split(',') if strs else LOG_OUTPUT_FORMATS
+    output_formats = [make_output_format(f, dir) for f in format_strs]
 
+    Logger.CURRENT = Logger(dir=dir, output_formats=output_formats)
+    log('Logging to %s'%dir)
 
-class session(object):
-    """
-    Context manager that sets up the loggers for an experiment.
-    """
-
-    CURRENT = None  # Set to a LoggerContext object using enter/exit or context manager
-
-    def __init__(self, dir=None, format_strs=None):
-        if dir is None:
-            dir = os.getenv('OPENAI_LOGDIR')
-        if dir is None:
-            dir = osp.join(tempfile.gettempdir(), 
-                datetime.datetime.now().strftime("openai-%Y-%m-%d-%H-%M-%S-%f"))
-        self.dir = dir
-        if format_strs is None:
-            format_strs = LOG_OUTPUT_FORMATS
-        output_formats = [make_output_format(f, dir) for f in format_strs]
-        Logger.CURRENT = Logger(dir=dir, output_formats=output_formats)
-        print('Logging to', dir)
-
-    def __enter__(self):
-        os.makedirs(self.evaluation_dir(), exist_ok=True)
-        output_formats = [make_output_format(f, self.evaluation_dir()) 
-                            for f in LOG_OUTPUT_FORMATS]
-        Logger.CURRENT = Logger(dir=self.dir, output_formats=output_formats)
-        os.environ['OPENAI_LOGDIR'] = self.evaluation_dir()
-
-    def __exit__(self, *args):
+def reset():
+    if Logger.CURRENT is not Logger.DEFAULT:
         Logger.CURRENT.close()
         Logger.CURRENT = Logger.DEFAULT
+        log('Reset logger')
 
-    def evaluation_dir(self):
-        return self.dir
-
-def _setup():
-    logdir = os.getenv('OPENAI_LOGDIR')
-    if logdir:
-        session(logdir).__enter__()
-
-_setup()
+class scoped_configure(object):
+    def __init__(self, dir=None, format_strs=None):
+        self.dir = dir
+        self.format_strs = format_strs
+        self.prevlogger = None
+    def __enter__(self):
+        self.prevlogger = Logger.CURRENT
+        configure(dir=self.dir, format_strs=self.format_strs)
+    def __exit__(self, *args):
+        Logger.CURRENT.close()
+        Logger.CURRENT = self.prevlogger
 
 # ================================================================
 
-
 def _demo():
     info("hi")
     debug("shouldn't appear")
@@ -318,21 +352,71 @@ def _demo():
     dir = "/tmp/testlogging"
     if os.path.exists(dir):
         shutil.rmtree(dir)
-    with session(dir=dir):
-        record_tabular("a", 3)
-        record_tabular("b", 2.5)
-        dump_tabular()
-        record_tabular("b", -2.5)
-        record_tabular("a", 5.5)
-        dump_tabular()
-        info("^^^ should see a = 5.5")
+    configure(dir=dir)
+    logkv("a", 3)
+    logkv("b", 2.5)
+    dumpkvs()
+    logkv("b", -2.5)
+    logkv("a", 5.5)
+    dumpkvs()
+    info("^^^ should see a = 5.5")
 
-    record_tabular("b", -2.5)
-    dump_tabular()
+    logkv("b", -2.5)
+    dumpkvs()
 
-    record_tabular("a", "longasslongasslongasslongasslongasslongassvalue")
-    dump_tabular()
+    logkv("a", "longasslongasslongasslongasslongasslongassvalue")
+    dumpkvs()
 
 
+# ================================================================
+# Readers
+# ================================================================
+
+def read_json(fname):
+    import pandas
+    ds = []
+    with open(fname, 'rt') as fh:
+        for line in fh:
+            ds.append(json.loads(line))
+    return pandas.DataFrame(ds)
+
+def read_csv(fname):
+    import pandas
+    return pandas.read_csv(fname, index_col=None, comment='#')
+
+def read_tb(path):
+    """
+    path : a tensorboard file OR a directory, where we will find all TB files
+           of the form events.*
+    """
+    import pandas
+    import numpy as np
+    from glob import glob
+    from collections import defaultdict
+    import tensorflow as tf
+    if osp.isdir(path):
+        fnames = glob(osp.join(path, "events.*"))
+    elif osp.basename(path).startswith("events."):
+        fnames = [path]
+    else:
+        raise NotImplementedError("Expected tensorboard file or directory containing them. Got %s"%path)
+    tag2pairs = defaultdict(list)
+    maxstep = 0
+    for fname in fnames:
+        for summary in tf.train.summary_iterator(fname):
+            if summary.step > 0:
+                for v in summary.summary.value:
+                    pair = (summary.step, v.simple_value)
+                    tag2pairs[v.tag].append(pair)
+                maxstep = max(summary.step, maxstep)
+    data = np.empty((maxstep, len(tag2pairs)))
+    data[:] = np.nan
+    tags = sorted(tag2pairs.keys())
+    for (colidx,tag) in enumerate(tags):
+        pairs = tag2pairs[tag]
+        for (step, value) in pairs:
+            data[step-1, colidx] = value
+    return pandas.DataFrame(data, columns=tags)
+
 if __name__ == "__main__":
     _demo()

baselines/ppo1/README.md

@@ -0,0 +1,7 @@
+# PPOSGD
+
+- Original paper: https://arxiv.org/abs/1707.06347
+- Baselines blog post: https://blog.openai.com/openai-baselines-ppo/
+- `mpirun -np 8 python -m baselines.ppo1.run_atari` runs the algorithm for 40M frames = 10M timesteps on an Atari game. See help (`-h`) for more options.
+- `python -m baselines.ppo1.run_mujoco` runs the algorithm for 1M frames on a Mujoco environment.
+

baselines/ppo1/cnn_policy.py

@@ -1,4 +1,3 @@
-from baselines.common.mpi_running_mean_std import RunningMeanStd
 import baselines.common.tf_util as U
 import tensorflow as tf
 import gym
@@ -49,7 +48,7 @@ class CnnPolicy(object):
         ac1, vpred1 =  self._act(stochastic, ob[None])
         return ac1[0], vpred1[0]
     def get_variables(self):
-        return tf.get_collection(tf.GraphKeys.VARIABLES, self.scope)
+        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):

baselines/ppo1/mlp_policy.py

@@ -51,7 +51,7 @@ class MlpPolicy(object):
         ac1, vpred1 =  self._act(stochastic, ob[None])
         return ac1[0], vpred1[0]
     def get_variables(self):
-        return tf.get_collection(tf.GraphKeys.VARIABLES, self.scope)
+        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):

baselines/ppo1/pposgd_simple.py

@@ -78,12 +78,13 @@ def add_vtarg_and_adv(seg, gamma, lam):
     seg["tdlamret"] = seg["adv"] + seg["vpred"]
 
 def learn(env, policy_func, *,
-        timesteps_per_batch, # timesteps per actor per update
+        timesteps_per_actorbatch, # timesteps per actor per update
         clip_param, entcoeff, # clipping parameter epsilon, entropy coeff
         optim_epochs, optim_stepsize, optim_batchsize,# optimization hypers
         gamma, lam, # advantage estimation
         max_timesteps=0, max_episodes=0, max_iters=0, max_seconds=0,  # time constraint
         callback=None, # you can do anything in the callback, since it takes locals(), globals()
+        adam_epsilon=1e-5,
         schedule='constant' # annealing for stepsize parameters (epsilon and adam)
         ):
     # Setup losses and stuff
@@ -111,17 +112,14 @@ def learn(env, policy_func, *,
     surr1 = ratio * atarg # surrogate from conservative policy iteration
     surr2 = U.clip(ratio, 1.0 - clip_param, 1.0 + clip_param) * atarg #
     pol_surr = - U.mean(tf.minimum(surr1, surr2)) # PPO's pessimistic surrogate (L^CLIP)
-    vfloss1 = tf.square(pi.vpred - ret)
-    vpredclipped = oldpi.vpred + tf.clip_by_value(pi.vpred - oldpi.vpred, -clip_param, clip_param)
-    vfloss2 = tf.square(vpredclipped - ret)
-    vf_loss = .5 * U.mean(tf.maximum(vfloss1, vfloss2)) # we do the same clipping-based trust region for the value function
+    vf_loss = U.mean(tf.square(pi.vpred - ret))
     total_loss = pol_surr + pol_entpen + vf_loss
     losses = [pol_surr, pol_entpen, vf_loss, meankl, meanent]
     loss_names = ["pol_surr", "pol_entpen", "vf_loss", "kl", "ent"]
 
     var_list = pi.get_trainable_variables()
     lossandgrad = U.function([ob, ac, atarg, ret, lrmult], losses + [U.flatgrad(total_loss, var_list)])
-    adam = MpiAdam(var_list)
+    adam = MpiAdam(var_list, epsilon=adam_epsilon)
 
     assign_old_eq_new = U.function([],[], updates=[tf.assign(oldv, newv)
         for (oldv, newv) in zipsame(oldpi.get_variables(), pi.get_variables())])
@@ -132,7 +130,7 @@ def learn(env, policy_func, *,
 
     # Prepare for rollouts
     # ----------------------------------------
-    seg_gen = traj_segment_generator(pi, env, timesteps_per_batch, stochastic=True)
+    seg_gen = traj_segment_generator(pi, env, timesteps_per_actorbatch, stochastic=True)
 
     episodes_so_far = 0
     timesteps_so_far = 0

baselines/ppo1/run_atari.py

@@ -0,0 +1,54 @@
+#!/usr/bin/env python
+
+from mpi4py import MPI
+from baselines.common import set_global_seeds
+from baselines import bench
+import os.path as osp
+import gym, logging
+from baselines import logger
+from baselines.common.atari_wrappers import make_atari, wrap_deepmind
+
+def train(env_id, num_timesteps, seed):
+    from baselines.ppo1 import pposgd_simple, cnn_policy
+    import baselines.common.tf_util as U
+    rank = MPI.COMM_WORLD.Get_rank()
+    sess = U.single_threaded_session()
+    sess.__enter__()
+    if rank == 0:
+        logger.configure()
+    else:
+        logger.configure(format_strs=[])
+    workerseed = seed + 10000 * MPI.COMM_WORLD.Get_rank()
+    set_global_seeds(workerseed)
+    env = make_atari(env_id)
+    def policy_fn(name, ob_space, ac_space): #pylint: disable=W0613
+        return cnn_policy.CnnPolicy(name=name, ob_space=ob_space, ac_space=ac_space)
+    env = bench.Monitor(env, logger.get_dir() and
+        osp.join(logger.get_dir(), str(rank)))
+    env.seed(workerseed)
+    gym.logger.setLevel(logging.WARN)
+
+    env = wrap_deepmind(env)
+    env.seed(workerseed)
+
+    pposgd_simple.learn(env, policy_fn,
+        max_timesteps=int(num_timesteps * 1.1),
+        timesteps_per_actorbatch=256,
+        clip_param=0.2, entcoeff=0.01,
+        optim_epochs=4, optim_stepsize=1e-3, optim_batchsize=64,
+        gamma=0.99, lam=0.95,
+        schedule='linear'
+    )
+    env.close()
+
+def main():
+    import argparse
+    parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter)
+    parser.add_argument('--env', help='environment ID', default='PongNoFrameskip-v4')
+    parser.add_argument('--seed', help='RNG seed', type=int, default=0)
+    parser.add_argument('--num-timesteps', type=int, default=int(10e6))
+    args = parser.parse_args()
+    train(args.env, num_timesteps=args.num_timesteps, seed=args.seed)
+
+if __name__ == '__main__':
+    main()

baselines/ppo1/run_mujoco.py

@@ -1,34 +1,38 @@
 #!/usr/bin/env python
 from baselines.common import set_global_seeds, tf_util as U
 from baselines import bench
-import os.path as osp
 import gym, logging
 from baselines import logger
-import sys
 
 def train(env_id, num_timesteps, seed):
-    from baselines.pposgd import mlp_policy, pposgd_simple
+    from baselines.ppo1 import mlp_policy, pposgd_simple
     U.make_session(num_cpu=1).__enter__()
-    logger.session().__enter__()
     set_global_seeds(seed)
     env = gym.make(env_id)
     def policy_fn(name, ob_space, ac_space):
         return mlp_policy.MlpPolicy(name=name, ob_space=ob_space, ac_space=ac_space,
             hid_size=64, num_hid_layers=2)
-    env = bench.Monitor(env, osp.join(logger.get_dir(), "monitor.json"))
+    env = bench.Monitor(env, logger.get_dir())
     env.seed(seed)
     gym.logger.setLevel(logging.WARN)
-    pposgd_simple.learn(env, policy_fn, 
+    pposgd_simple.learn(env, policy_fn,
             max_timesteps=num_timesteps,
-            timesteps_per_batch=2048,
+            timesteps_per_actorbatch=2048,
             clip_param=0.2, entcoeff=0.0,
             optim_epochs=10, optim_stepsize=3e-4, optim_batchsize=64,
-            gamma=0.99, lam=0.95,
+            gamma=0.99, lam=0.95, schedule='linear',
         )
     env.close()
 
 def main():
-    train('Hopper-v1', num_timesteps=1e6, seed=0)
+    import argparse
+    parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter)
+    parser.add_argument('--env', help='environment ID', default='Hopper-v1')
+    parser.add_argument('--seed', help='RNG seed', type=int, default=0)
+    parser.add_argument('--num-timesteps', type=int, default=int(1e6))
+    args = parser.parse_args()
+    logger.configure()
+    train(args.env, num_timesteps=args.num_timesteps, seed=args.seed)
 
 
 if __name__ == '__main__':

baselines/pposgd/run_atari.py

@@ -1,54 +0,0 @@
-#!/usr/bin/env python
-from mpi4py import MPI
-from baselines.common import set_global_seeds
-from baselines import bench
-from baselines.common.mpi_fork import mpi_fork
-import os.path as osp
-import gym, logging
-from baselines import logger
-import sys
-
-def wrap_train(env):
-    from baselines.common.atari_wrappers import (wrap_deepmind, FrameStack)
-    env = wrap_deepmind(env, clip_rewards=True)
-    env = FrameStack(env, 4)
-    return env
-
-def train(env_id, num_timesteps, seed, num_cpu):
-    from baselines.pposgd import pposgd_simple, cnn_policy
-    import baselines.common.tf_util as U
-    whoami  = mpi_fork(num_cpu)
-    if whoami == "parent": return
-    rank = MPI.COMM_WORLD.Get_rank()
-    sess = U.single_threaded_session()
-    sess.__enter__()
-    logger.session().__enter__()
-    if rank != 0: logger.set_level(logger.DISABLED)
-    workerseed = seed + 10000 * MPI.COMM_WORLD.Get_rank()
-    set_global_seeds(workerseed)
-    env = gym.make(env_id)
-    def policy_fn(name, ob_space, ac_space): #pylint: disable=W0613
-        return cnn_policy.CnnPolicy(name=name, ob_space=ob_space, ac_space=ac_space)
-    env = bench.Monitor(env, osp.join(logger.get_dir(), "%i.monitor.json" % rank))
-    env.seed(workerseed)
-    gym.logger.setLevel(logging.WARN)
-
-    env = wrap_train(env)
-    num_timesteps /= 4 # because we're wrapping the envs to do frame skip
-    env.seed(workerseed)
-
-    pposgd_simple.learn(env, policy_fn,
-        max_timesteps=num_timesteps,
-        timesteps_per_batch=256,
-        clip_param=0.2, entcoeff=0.01,
-        optim_epochs=4, optim_stepsize=1e-3, optim_batchsize=64,
-        gamma=0.99, lam=0.95,
-        schedule='linear'
-    )
-    env.close()
-
-def main():
-    train('PongNoFrameskip-v4', num_timesteps=40e6, seed=0, num_cpu=8)
-
-if __name__ == '__main__':
-    main()

baselines/trpo_mpi/README.md

@@ -0,0 +1,6 @@
+# trpo_mpi
+
+- Original paper: https://arxiv.org/abs/1502.05477
+- Baselines blog post https://blog.openai.com/openai-baselines-ppo/
+- `mpirun -np 16 python -m baselines.trpo_mpi.run_atari` runs the algorithm for 40M frames = 10M timesteps on an Atari game. See help (`-h`) for more options.
+- `python -m baselines.trpo_mpi.run_mujoco` runs the algorithm for 1M timesteps on a Mujoco environment.

baselines/trpo_mpi/nosharing_cnn_policy.py

@@ -1,4 +1,3 @@
-from baselines.common.mpi_running_mean_std import RunningMeanStd
 import baselines.common.tf_util as U
 import tensorflow as tf
 import gym
@@ -42,14 +41,14 @@ class CnnPolicy(object):
         self.state_out = []
 
         stochastic = tf.placeholder(dtype=tf.bool, shape=())
-        ac = self.pd.sample() # XXX
+        ac = self.pd.sample()
         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.VARIABLES, self.scope)
+        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):

baselines/trpo_mpi/run_atari.py

@@ -1,53 +1,49 @@
-#!/usr/bin/env python
+    #!/usr/bin/env python
 from mpi4py import MPI
 from baselines.common import set_global_seeds
 import os.path as osp
 import gym, logging
 from baselines import logger
 from baselines import bench
-from baselines.common.mpi_fork import mpi_fork
-import sys
+from baselines.common.atari_wrappers import make_atari, wrap_deepmind
 
-def wrap_train(env):
-    from baselines.common.atari_wrappers import (wrap_deepmind, FrameStack)
-    env = wrap_deepmind(env, clip_rewards=False)
-    env = FrameStack(env, 3)
-    return env
-
-def train(env_id, num_timesteps, seed, num_cpu):
+def train(env_id, num_timesteps, seed):
     from baselines.trpo_mpi.nosharing_cnn_policy import CnnPolicy
     from baselines.trpo_mpi import trpo_mpi
     import baselines.common.tf_util as U
-    whoami  = mpi_fork(num_cpu)
-    if whoami == "parent":
-        return
     rank = MPI.COMM_WORLD.Get_rank()
     sess = U.single_threaded_session()
     sess.__enter__()
-    logger.session().__enter__()
-    if rank != 0:
-        logger.set_level(logger.DISABLED)
-
+    if rank == 0:
+        logger.configure()
+    else:
+        logger.configure(format_strs=[])
 
     workerseed = seed + 10000 * MPI.COMM_WORLD.Get_rank()
     set_global_seeds(workerseed)
-    env = gym.make(env_id)
+    env = make_atari(env_id)
     def policy_fn(name, ob_space, ac_space): #pylint: disable=W0613
         return CnnPolicy(name=name, ob_space=env.observation_space, ac_space=env.action_space)
-    env = bench.Monitor(env, osp.join(logger.get_dir(), "%i.monitor.json"%rank))
+    env = bench.Monitor(env, logger.get_dir() and osp.join(logger.get_dir(), str(rank)))
     env.seed(workerseed)
     gym.logger.setLevel(logging.WARN)
 
-    env = wrap_train(env)
-    num_timesteps /= 4 # because we're wrapping the envs to do frame skip
+    env = wrap_deepmind(env)
     env.seed(workerseed)
 
     trpo_mpi.learn(env, policy_fn, timesteps_per_batch=512, max_kl=0.001, cg_iters=10, cg_damping=1e-3,
-        max_timesteps=num_timesteps, gamma=0.98, lam=1.0, vf_iters=3, vf_stepsize=1e-4, entcoeff=0.00)
+        max_timesteps=int(num_timesteps * 1.1), gamma=0.98, lam=1.0, vf_iters=3, vf_stepsize=1e-4, entcoeff=0.00)
     env.close()
 
 def main():
-    train('PongNoFrameskip-v4', num_timesteps=40e6, seed=0, num_cpu=8)
+    import argparse
+    parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter)
+    parser.add_argument('--env', help='environment ID', default='PongNoFrameskip-v4')
+    parser.add_argument('--seed', help='RNG seed', type=int, default=0)
+    parser.add_argument('--num-timesteps', type=int, default=int(10e6))
+    args = parser.parse_args()
+    train(args.env, num_timesteps=args.num_timesteps, seed=args.seed)
+
 
 if __name__ == "__main__":
     main()

baselines/trpo_mpi/run_mujoco.py

@@ -7,19 +7,14 @@ import os.path as osp
 import gym
 import logging
 from baselines import logger
-from baselines.pposgd.mlp_policy import MlpPolicy
+from baselines.ppo1.mlp_policy import MlpPolicy
 from baselines.common.mpi_fork import mpi_fork
 from baselines import bench
 from baselines.trpo_mpi import trpo_mpi
 import sys
-num_cpu=1
 
 def train(env_id, num_timesteps, seed):
-    whoami  = mpi_fork(num_cpu)
-    if whoami == "parent":
-        return
     import baselines.common.tf_util as U
-    logger.session().__enter__()
     sess = U.single_threaded_session()
     sess.__enter__()
 
@@ -32,7 +27,8 @@ def train(env_id, num_timesteps, seed):
     def policy_fn(name, ob_space, ac_space):
         return MlpPolicy(name=name, ob_space=env.observation_space, ac_space=env.action_space,
             hid_size=32, num_hid_layers=2)
-    env = bench.Monitor(env, osp.join(logger.get_dir(), "%i.monitor.json" % rank))
+    env = bench.Monitor(env, logger.get_dir() and
+        osp.join(logger.get_dir(), str(rank)))
     env.seed(workerseed)
     gym.logger.setLevel(logging.WARN)
 
@@ -41,7 +37,15 @@ def train(env_id, num_timesteps, seed):
     env.close()
 
 def main():
-    train('Hopper-v1', num_timesteps=1e6, seed=0)
+    import argparse
+    parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter)
+    parser.add_argument('--env', help='environment ID', default='Hopper-v1')
+    parser.add_argument('--seed', help='RNG seed', type=int, default=0)
+    parser.add_argument('--num-timesteps', type=int, default=int(1e6))
+    args = parser.parse_args()
+    logger.configure()
+    train(args.env, num_timesteps=args.num_timesteps, seed=args.seed)
+
 
 if __name__ == '__main__':
     main()

baselines/trpo_mpi/trpo_mpi.py

@@ -207,7 +207,7 @@ def learn(env, policy_func, *,
         if hasattr(pi, "ret_rms"): pi.ret_rms.update(tdlamret)
         if hasattr(pi, "ob_rms"): pi.ob_rms.update(ob) # update running mean/std for policy
 
-        args = seg["ob"], seg["ac"], seg["adv"]
+        args = seg["ob"], seg["ac"], atarg
         fvpargs = [arr[::5] for arr in args]
         def fisher_vector_product(p):
             return allmean(compute_fvp(p, *fvpargs)) + cg_damping * p

setup.py

@@ -10,7 +10,7 @@ setup(name='baselines',
       packages=[package for package in find_packages()
                 if package.startswith('baselines')],
       install_requires=[
-          'gym>=0.9.1[all]',
+          'gym[mujoco,atari,classic_control]',
           'scipy',
           'tqdm',
           'joblib',
@@ -19,6 +19,7 @@ setup(name='baselines',
           'tensorflow >= 1.0.0',
           'azure==1.0.3',
           'progressbar2',
+          'mpi4py',
       ],
       description="OpenAI baselines: high quality implementations of reinforcement learning algorithms",
       author="OpenAI",