RNN support for PPO2 (#859)

* initial implementaion of ppo2_rnn.

* set lstm memory as tf.GraphKeys.LOCAL_VARIABLES.

* replace dones with tf.placeholder_with_default.

* improves for 'play' option.

* removed unnecessary TODO .

* improve lstm code.

* move learning rate placeholer to optimizer scope.

* support the microbatched model.

* sync cnn lstm layer with originals.

* add cnn_lnlstm layer.

* fix a case when `states` is None.

* add initial_state variable to help test.

* make ppo2 rnn test available.

* rename 'obs' with 'observations'.
rename 'transition' with 'transitions'.
fix forgetting `dones` in the replay buffer.
fix a misuse of `states` and `next_states` in the replay buffer.

* make initialization once.
make `test_fixed_sequence` compatible with ppo2.

* adjust input shape.

* fix checking of a model input args in `simple_test` function.

* disable warning on purpose.

* support the play.

* improve scopes to compatible with multiple models (i.e, other tensorflow global/local variables)

* clean the scope of ppo2 policy model.

* name the memory variable of PPO RNNs more describly

* wrap the initializations in ppo2.

* remove redundant lines.

* update `REAMD.md`.

* add RNN layers.

* add the result of HalfCheeta-v2 env  experiment.

* correct a typo.

* add RNN class.

* rename `nlstm` with `num_units` in RNN builder functions.

* remove state saving.

* reuse RNNs in a2c.utils.

* revert baselines/run.py.

* replace `ppo2.step()` with original interface.

* revert `baselines/common/tests/util.py`.

* remove redundant lines.

* revert `baselines/common/test/util.py` to b875fb7.

* remove `states` variable.

* move RNN class to `baselines/ppo2/layers.py' and revert `baselines/common/models.py` to 858afa8.

* rename `model.step_as_dict` with `model.step_with_dict`.

* removed `ppo_lstm_mlp`.

* fix 02e26fd.

baselines/common/tests/test_fixed_sequence.py

@@ -17,10 +17,10 @@ learn_kwargs = {
     # 'trpo_mpi': lambda e, p: trpo_mpi.learn(policy_fn=p(env=e), env=e, max_timesteps=30000, timesteps_per_batch=100, cg_iters=10, gamma=0.9, lam=1.0, max_kl=0.001)
 }
 
-
 alg_list = learn_kwargs.keys()
 rnn_list = ['lstm']
 
+
 @pytest.mark.slow
 @pytest.mark.parametrize("alg", alg_list)
 @pytest.mark.parametrize("rnn", rnn_list)
@@ -33,6 +33,9 @@ def test_fixed_sequence(alg, rnn):
     kwargs = learn_kwargs[alg]
     kwargs.update(common_kwargs)
 
+    if alg == 'ppo2' and rnn.endswith('lstm'):
+        rnn = 'ppo_' + rnn
+
     env_fn = lambda: FixedSequenceEnv(n_actions=10, episode_len=5)
     learn = lambda e: get_learn_function(alg)(
         env=e,
@@ -45,6 +48,3 @@ def test_fixed_sequence(alg, rnn):
 
 if __name__ == '__main__':
     test_fixed_sequence('ppo2', 'lstm')
-
-
-

baselines/common/tests/test_serialization.py

@@ -1,17 +1,16 @@
 import os
-import gym
 import tempfile
-import pytest
-import tensorflow as tf
-import numpy as np
-
-from baselines.common.tests.envs.mnist_env import MnistEnv
-from baselines.common.vec_env.dummy_vec_env import DummyVecEnv
-from baselines.run import get_learn_function
-from baselines.common.tf_util import make_session, get_session
-
 from functools import partial
 
+import gym
+import numpy as np
+import pytest
+import tensorflow as tf
+
+from baselines.common.tests.envs.mnist_env import MnistEnv
+from baselines.common.tf_util import make_session, get_session
+from baselines.common.vec_env.dummy_vec_env import DummyVecEnv
+from baselines.run import get_learn_function
 
 learn_kwargs = {
     'deepq': {},
@@ -37,12 +36,15 @@ def test_serialization(learn_fn, network_fn):
     Test if the trained model can be serialized
     '''
 
+    _network_kwargs = network_kwargs[network_fn]
 
     if network_fn.endswith('lstm') and learn_fn in ['acer', 'acktr', 'trpo_mpi', 'deepq']:
         # TODO make acktr work with recurrent policies
         # and test
         # github issue: https://github.com/openai/baselines/issues/660
         return
+    elif network_fn.endswith('lstm') and learn_fn == 'ppo2':
+        network_fn = 'ppo_' + network_fn
 
     def make_env():
         env = MnistEnv(episode_len=100)
@@ -54,10 +56,9 @@ def test_serialization(learn_fn, network_fn):
     learn = get_learn_function(learn_fn)
 
     kwargs = {}
-    kwargs.update(network_kwargs[network_fn])
+    kwargs.update(_network_kwargs)
     kwargs.update(learn_kwargs[learn_fn])
 
-
     learn = partial(learn, env=env, network=network_fn, seed=0, **kwargs)
 
     with tempfile.TemporaryDirectory() as td:
@@ -117,7 +118,6 @@ def test_coexistence(learn_fn, network_fn):
     model2.step(env.observation_space.sample())
 
 
-
 def _serialize_variables():
     sess = get_session()
     variables = tf.trainable_variables()
@@ -136,4 +136,3 @@ def _get_action_stats(model, ob):
     std = np.std(actions, axis=0)
 
     return mean, std
-

baselines/ppo2/README.md

@@ -3,6 +3,16 @@
 - Original paper: https://arxiv.org/abs/1707.06347
 - Baselines blog post: https://blog.openai.com/openai-baselines-ppo/
 
+## Examples 
 - `python -m baselines.run --alg=ppo2 --env=PongNoFrameskip-v4` runs the algorithm for 40M frames = 10M timesteps on an Atari Pong. See help (`-h`) for more options.
 - `python -m baselines.run --alg=ppo2 --env=Ant-v2 --num_timesteps=1e6` runs the algorithm for 1M frames on a Mujoco Ant environment.
-- also refer to the repo-wide [README.md](../../README.md#training-models)
+
+### RNN networks
+- `python -m baselines.run --alg=ppo2 --env=PongNoFrameskip-v4 --network=ppo_cnn_lstm` runs on an Atari Pong with 
+    `ppo_cnn_lstm` network.
+- `python -m baselines.run --alg=ppo2 --env=Ant-v2 --num_timesteps=1e6 --network=ppo_lstm --value_network=copy` 
+    runs on a Mujoco Ant environment with `ppo_lstm` network whose value and policy networks are separated, but have 
+    same structure.
+
+## See Also
+- refer to the repo-wide [README.md](../../README.md#training-models)

baselines/ppo2/__init__.py

@@ -0,0 +1 @@
+from baselines.ppo2.layers import ppo_lstm, ppo_cnn_lstm, ppo_cnn_lnlstm  # pylint: disable=unused-import # noqa: F401

baselines/ppo2/layers.py

@@ -0,0 +1,55 @@
+import numpy as np
+import tensorflow as tf
+
+from baselines.a2c.utils import ortho_init, lstm, lnlstm
+from baselines.common.models import register, nature_cnn
+
+
+class RNN(object):
+    def __init__(self, func, memory_size=None):
+        self._func = func
+        self.memory_size = memory_size
+
+    def __call__(self, *args, **kwargs):
+        return self._func(*args, **kwargs)
+
+
+@register("ppo_lstm")
+def ppo_lstm(num_units=128, layer_norm=False):
+    def network_fn(input, mask, state):
+        input = tf.layers.flatten(input)
+        mask = tf.to_float(mask)
+
+        if layer_norm:
+            h, next_state = lnlstm([input], [mask[:, None]], state, scope='lnlstm', nh=num_units)
+        else:
+            h, next_state = lstm([input], [mask[:, None]], state, scope='lstm', nh=num_units)
+        h = h[0]
+        return h, next_state
+
+    return RNN(network_fn, memory_size=num_units * 2)
+
+
+@register("ppo_cnn_lstm")
+def ppo_cnn_lstm(num_units=128, layer_norm=False, **conv_kwargs):
+    def network_fn(input, mask, state):
+        mask = tf.to_float(mask)
+        initializer = ortho_init(np.sqrt(2))
+
+        h = nature_cnn(input, **conv_kwargs)
+        h = tf.layers.flatten(h)
+        h = tf.layers.dense(h, units=512, activation=tf.nn.relu, kernel_initializer=initializer)
+
+        if layer_norm:
+            h, next_state = lnlstm([h], [mask[:, None]], state, scope='lnlstm', nh=num_units)
+        else:
+            h, next_state = lstm([h], [mask[:, None]], state, scope='lstm', nh=num_units)
+        h = h[0]
+        return h, next_state
+
+    return RNN(network_fn, memory_size=num_units * 2)
+
+
+@register("ppo_cnn_lnlstm")
+def ppo_cnn_lnlstm(num_units=128, **conv_kwargs):
+    return ppo_cnn_lstm(num_units, layer_norm=True, **conv_kwargs)

baselines/ppo2/microbatched_model.py

@@ -1,18 +1,21 @@
-import tensorflow as tf
 import numpy as np
+import tensorflow as tf
 from baselines.ppo2.model import Model
 
+
 class MicrobatchedModel(Model):
     """
     Model that does training one microbatch at a time - when gradient computation
     on the entire minibatch causes some overflow
     """
+
     def __init__(self, *, policy, ob_space, ac_space, nbatch_act, nbatch_train,
                  nsteps, ent_coef, vf_coef, max_grad_norm, microbatch_size):
 
         self.nmicrobatches = nbatch_train // microbatch_size
         self.microbatch_size = microbatch_size
-        assert nbatch_train % microbatch_size == 0, 'microbatch_size ({}) should divide nbatch_train ({}) evenly'.format(microbatch_size, nbatch_train)
+        assert nbatch_train % microbatch_size == 0, 'microbatch_size ({}) should divide nbatch_train ({}) evenly'.format(
+            microbatch_size, nbatch_train)
 
         super().__init__(
             policy=policy,
@@ -29,14 +32,16 @@ class MicrobatchedModel(Model):
         grads_ph_and_vars = list(zip(self.grads_ph, self.var))
         self._apply_gradients_op = self.trainer.apply_gradients(grads_ph_and_vars)
 
-
+    def train(self,
-    def train(self, lr, cliprange, obs, returns, masks, actions, values, neglogpacs, states=None):
+              lr,
-        assert states is None, "microbatches with recurrent models are not supported yet"
+              cliprange,
-
+              observations,
-        # Here we calculate advantage A(s,a) = R + yV(s') - V(s)
+              advs,
-        # Returns = R + yV(s')
+              returns,
-        advs = returns - values
+              actions,
-
+              values,
+              neglogpacs,
+              **_kwargs):
         # Normalize the advantages
         advs = (advs - advs.mean()) / (advs.std() + 1e-8)
 
@@ -45,16 +50,21 @@ class MicrobatchedModel(Model):
 
         for microbatch_idx in range(self.nmicrobatches):
             _sli = range(microbatch_idx * self.microbatch_size, (microbatch_idx + 1) * self.microbatch_size)
+
             td_map = {
-                self.train_model.X: obs[_sli],
+                self.train_model.X: observations[_sli],
                 self.A: actions[_sli],
                 self.ADV: advs[_sli],
-                self.R:returns[_sli],
+                self.RETURNS: returns[_sli],
+                self.LR: lr,
                 self.CLIPRANGE: cliprange,
                 self.OLDNEGLOGPAC: neglogpacs[_sli],
-                self.OLDVPRED:values[_sli]
+                self.VALUE_PREV: values[_sli],
             }
 
+            sliced_kwargs = {key: _kwargs[key][_sli] for key in _kwargs}
+            td_map.update(self.train_model.feed_dict(**sliced_kwargs))
+
             # Compute gradient on a microbatch (note that variables do not change here) ...
             grad_v, stats_v = self.sess.run([self.grads, self.stats_list], td_map)
             if microbatch_idx == 0:
@@ -71,6 +81,3 @@ class MicrobatchedModel(Model):
         self.sess.run(self._apply_gradients_op, feed_dict)
         # Return average of the stats
         return np.mean(np.array(stats_vs), axis=0).tolist()
-
-
-

baselines/ppo2/model.py

@@ -1,8 +1,8 @@
-import tensorflow as tf
 import functools
 
+import tensorflow as tf
+
 from baselines.common.tf_util import get_session, save_variables, load_variables
-from baselines.common.tf_util import initialize
 
 try:
     from baselines.common.mpi_adam_optimizer import MpiAdamOptimizer
@@ -11,6 +11,7 @@ try:
 except ImportError:
     MPI = None
 
+
 class Model(object):
     """
     We use this object to :
@@ -24,72 +25,82 @@ class Model(object):
     save/load():
     - Save load the model
     """
-    def __init__(self, *, policy, ob_space, ac_space, nbatch_act, nbatch_train,
-                nsteps, ent_coef, vf_coef, max_grad_norm, microbatch_size=None):
-        self.sess = sess = get_session()
 
-        with tf.variable_scope('ppo2_model', reuse=tf.AUTO_REUSE):
+    def __init__(self, *, policy, ob_space, ac_space, nbatch_act, nbatch_train,
+                 nsteps, ent_coef, vf_coef, max_grad_norm,
+                 name='ppo_model',
+                 sess=None,
+                 microbatch_size=None):
+        if sess is None:
+            sess = get_session()
+        self.sess = sess
+        self.name = name
+
+        with tf.variable_scope(name) as scope:
+            self.scope = scope
+            with tf.variable_scope('models', reuse=tf.AUTO_REUSE):
+                with tf.name_scope('act_model'):
                     # CREATE OUR TWO MODELS
                     # act_model that is used for sampling
                     act_model = policy(nbatch_act, 1, sess)
 
+                with tf.name_scope('train_model'):
                     # Train model for training
                     if microbatch_size is None:
                         train_model = policy(nbatch_train, nsteps, sess)
                     else:
                         train_model = policy(microbatch_size, nsteps, sess)
 
+            with tf.variable_scope('losses'):
                 # CREATE THE PLACEHOLDERS
-        self.A = A = train_model.pdtype.sample_placeholder([None])
+                self.A = A = train_model.pdtype.sample_placeholder([None], name='action')
-        self.ADV = ADV = tf.placeholder(tf.float32, [None])
+                self.ADV = ADV = tf.placeholder(tf.float32, [None], name='advantage')
-        self.R = R = tf.placeholder(tf.float32, [None])
+                self.RETURNS = RETURNS = tf.placeholder(tf.float32, [None], name='reward')
-        # Keep track of old actor
+                self.VALUE_PREV = VALUE_PREV = tf.placeholder(tf.float32, [None], name='value_prev')
-        self.OLDNEGLOGPAC = OLDNEGLOGPAC = tf.placeholder(tf.float32, [None])
+                self.OLDNEGLOGPAC = OLDNEGLOGPAC = tf.placeholder(tf.float32, [None],
-        # Keep track of old critic
+                                                                  name='negative_log_p_action_old')
-        self.OLDVPRED = OLDVPRED = tf.placeholder(tf.float32, [None])
+                self.CLIPRANGE = CLIPRANGE = tf.placeholder(tf.float32, [], name='clip_range')
-        self.LR = LR = tf.placeholder(tf.float32, [])
-        # Cliprange
-        self.CLIPRANGE = CLIPRANGE = tf.placeholder(tf.float32, [])
 
+                with tf.name_scope("neglogpac"):
                     neglogpac = train_model.pd.neglogp(A)
 
+                with tf.name_scope("entropy"):
                     # Calculate the entropy
                     # Entropy is used to improve exploration by limiting the premature convergence to suboptimal policy.
                     entropy = tf.reduce_mean(train_model.pd.entropy())
+                    entropy_loss = (- ent_coef) * entropy
 
+                with tf.name_scope("value_loss"):
                     # CALCULATE THE LOSS
-        # Total loss = Policy gradient loss - entropy * entropy coefficient + Value coefficient * value loss
+                    value = train_model.value
-
+                    value_clipped = VALUE_PREV + tf.clip_by_value(value - VALUE_PREV, -CLIPRANGE, CLIPRANGE)
-        # Clip the value to reduce variability during Critic training
+                    vf_losses1 = tf.squared_difference(value, RETURNS)
-        # Get the predicted value
+                    vf_losses2 = tf.squared_difference(value_clipped, RETURNS)
-        vpred = train_model.vf
+                    vf_loss = 0.5 * vf_coef * tf.reduce_mean(tf.maximum(vf_losses1, vf_losses2))
-        vpredclipped = OLDVPRED + tf.clip_by_value(train_model.vf - OLDVPRED, - CLIPRANGE, CLIPRANGE)
-        # Unclipped value
-        vf_losses1 = tf.square(vpred - R)
-        # Clipped value
-        vf_losses2 = tf.square(vpredclipped - R)
-
-        vf_loss = .5 * tf.reduce_mean(tf.maximum(vf_losses1, vf_losses2))
 
+                with tf.name_scope("policy_loss"):
                     # Calculate ratio (pi current policy / pi old policy)
                     ratio = tf.exp(OLDNEGLOGPAC - neglogpac)
-
-        # Defining Loss = - J is equivalent to max J
                     pg_losses = -ADV * ratio
-
                     pg_losses2 = -ADV * tf.clip_by_value(ratio, 1.0 - CLIPRANGE, 1.0 + CLIPRANGE)
-
-        # Final PG loss
                     pg_loss = tf.reduce_mean(tf.maximum(pg_losses, pg_losses2))
-        approxkl = .5 * tf.reduce_mean(tf.square(neglogpac - OLDNEGLOGPAC))
+
+                with tf.name_scope("approxkl"):
+                    approxkl = .5 * tf.reduce_mean(tf.squared_difference(neglogpac, OLDNEGLOGPAC))
+
+                with tf.name_scope("clip_fraction"):
                     clipfrac = tf.reduce_mean(tf.to_float(tf.greater(tf.abs(ratio - 1.0), CLIPRANGE)))
 
-        # Total loss
+                with tf.name_scope("total_loss"):
-        loss = pg_loss - entropy * ent_coef + vf_loss * vf_coef
+                    loss = pg_loss + entropy_loss + vf_loss
+
+            with tf.variable_scope('optimizer'):
+                self.LR = LR = tf.placeholder(tf.float32, [], name='learning_rate')
 
                 # UPDATE THE PARAMETERS USING LOSS
                 # 1. Get the model parameters
-        params = tf.trainable_variables('ppo2_model')
+                params = tf.trainable_variables(self.scope.name)
+
                 # 2. Build our trainer
                 if MPI is not None:
                     self.trainer = MpiAdamOptimizer(MPI.COMM_WORLD, learning_rate=LR, epsilon=1e-5)
@@ -103,54 +114,68 @@ class Model(object):
                     # Clip the gradients (normalize)
                     grads, _grad_norm = tf.clip_by_global_norm(grads, max_grad_norm)
                 grads_and_var = list(zip(grads, var))
-        # zip aggregate each gradient with parameters associated
-        # For instance zip(ABCD, xyza) => Ax, By, Cz, Da
 
                 self.grads = grads
                 self.var = var
                 self._train_op = self.trainer.apply_gradients(grads_and_var)
-        self.loss_names = ['policy_loss', 'value_loss', 'policy_entropy', 'approxkl', 'clipfrac']
-        self.stats_list = [pg_loss, vf_loss, entropy, approxkl, clipfrac]
 
+                self.loss_names = ['policy_loss', 'value_loss', 'entropy_loss', 'approxkl', 'clipfrac',
+                                   'total_loss']
+                self.stats_list = [pg_loss, vf_loss, entropy_loss, approxkl, clipfrac, loss]
 
                 self.train_model = train_model
                 self.act_model = act_model
-        self.step = act_model.step
-        self.value = act_model.value
                 self.initial_state = act_model.initial_state
 
                 self.save = functools.partial(save_variables, sess=sess)
                 self.load = functools.partial(load_variables, sess=sess)
 
-        initialize()
+            with tf.variable_scope('initialization'):
-        global_variables = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope="")
+                sess.run(tf.initializers.variables(tf.global_variables(self.scope.name)))
+                sess.run(tf.initializers.variables(tf.local_variables(self.scope.name)))
+                global_variables = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope=self.scope.name)
                 if MPI is not None:
                     sync_from_root(sess, global_variables)  # pylint: disable=E1101
 
-    def train(self, lr, cliprange, obs, returns, masks, actions, values, neglogpacs, states=None):
+    def step_with_dict(self, **kwargs):
-        # Here we calculate advantage A(s,a) = R + yV(s') - V(s)
+        return self.act_model.step(**kwargs)
-        # Returns = R + yV(s')
-        advs = returns - values
 
+    def step(self, obs, M=None, S=None, **kwargs):
+        kwargs.update({'observations': obs})
+        if M is not None and S is not None:
+            kwargs.update({'dones': M})
+            kwargs.update({'states': S})
+        transition = self.act_model.step(**kwargs)
+        states = transition['next_states'] if 'next_states' in transition else None
+        return transition['actions'], transition['values'], states, transition['neglogpacs']
+
+    def train(self,
+              lr,
+              cliprange,
+              observations,
+              advs,
+              returns,
+              actions,
+              values,
+              neglogpacs,
+              **_kwargs):
         # Normalize the advantages
         advs = (advs - advs.mean()) / (advs.std() + 1e-8)
 
         td_map = {
-            self.train_model.X : obs,
+            self.train_model.X: observations,
             self.A: actions,
             self.ADV: advs,
-            self.R : returns,
+            self.RETURNS: returns,
             self.LR: lr,
             self.CLIPRANGE: cliprange,
             self.OLDNEGLOGPAC: neglogpacs,
-            self.OLDVPRED : values
+            self.VALUE_PREV: values,
         }
-        if states is not None:
+
-            td_map[self.train_model.S] = states
+        td_map.update(self.train_model.feed_dict(**_kwargs))
-            td_map[self.train_model.M] = masks
 
         return self.sess.run(
             self.stats_list + [self._train_op],
             td_map
         )[:-1]
-

baselines/ppo2/policies.py

@@ -0,0 +1,188 @@
+import gym
+import numpy as np
+import tensorflow as tf
+
+from baselines.a2c.utils import fc
+from baselines.common import tf_util
+from baselines.common.distributions import make_pdtype
+from baselines.common.input import observation_placeholder, encode_observation
+from baselines.common.models import get_network_builder
+from baselines.common.tf_util import adjust_shape
+from baselines.ppo2.layers import RNN
+
+
+class PolicyWithValue(object):
+    """
+    Encapsulates fields and methods for RL policy and two value function estimation with shared parameters
+    """
+
+    def __init__(self, env, observations, latent, dones, states=None, estimate_q=False, vf_latent=None, sess=None):
+        """
+        Parameters:
+        ----------
+        env             RL environment
+
+        observations    tensorflow placeholder in which the observations will be fed
+
+        latent          latent state from which policy distribution parameters should be inferred
+
+        vf_latent       latent state from which value function should be inferred (if None, then latent is used)
+
+        sess            tensorflow session to run calculations in (if None, default session is used)
+
+        **tensors       tensorflow tensors for additional attributes such as state or mask
+
+        """
+        self.X = observations
+        self.dones = dones
+        self.pdtype = make_pdtype(env.action_space)
+        self.states = states
+        self.sess = sess or tf.get_default_session()
+
+        vf_latent = vf_latent if vf_latent is not None else latent
+
+        with tf.variable_scope('policy'):
+            latent = tf.layers.flatten(latent)
+            # Based on the action space, will select what probability distribution type
+            self.pd, self.pi = self.pdtype.pdfromlatent(latent, init_scale=0.01)
+
+            with tf.variable_scope('sample_action'):
+                self.action = self.pd.sample()
+
+            with tf.variable_scope('negative_log_probability'):
+                # Calculate the neg log of our probability
+                self.neglogp = self.pd.neglogp(self.action)
+
+        with tf.variable_scope('value'):
+            vf_latent = tf.layers.flatten(vf_latent)
+
+            if estimate_q:
+                assert isinstance(env.action_space, gym.spaces.Discrete)
+                self.q = fc(vf_latent, 'q', env.action_space.n)
+                self.value = self.q
+            else:
+                vf_latent = tf.layers.flatten(vf_latent)
+                self.value = fc(vf_latent, 'value', 1, init_scale=0.01)
+                self.value = self.value[:, 0]
+
+        self.step_input = {
+            'observations': observations,
+            'dones': self.dones,
+        }
+
+        self.step_output = {
+            'actions': self.action,
+            'values': self.value,
+            'neglogpacs': self.neglogp,
+        }
+        if self.states:
+            self.initial_state = np.zeros(self.states['current'].get_shape())
+            self.step_input.update({'states': self.states['current']})
+            self.step_output.update({'states': self.states['current'],
+                                     'next_states': self.states['next']})
+        else:
+            self.initial_state = None
+
+    def feed_dict(self, **kwargs):
+        feed_dict = {}
+        for key in kwargs:
+            if key in self.step_input:
+                feed_dict[self.step_input[key]] = adjust_shape(self.step_input[key], kwargs[key])
+        return feed_dict
+
+    def step(self, **kwargs):
+        return self.sess.run(self.step_output,
+                             feed_dict=self.feed_dict(**kwargs))
+
+    def values(self, **kwargs):
+        return self.sess.run({'values': self.value},
+                             feed_dict=self.feed_dict(**kwargs))
+
+    def save(self, save_path):
+        tf_util.save_state(save_path, sess=self.sess)
+
+    def load(self, load_path):
+        tf_util.load_state(load_path, sess=self.sess)
+
+
+def build_ppo_policy(env, policy_network, value_network=None, estimate_q=False, **policy_kwargs):
+    if isinstance(policy_network, str):
+        network_type = policy_network
+        policy_network = get_network_builder(network_type)(**policy_kwargs)
+
+    if value_network is None:
+        value_network = 'shared'
+
+    def policy_fn(nbatch=None, nsteps=None, sess=None, observ_placeholder=None):
+        next_states_list = []
+        state_map = {}
+        state_placeholder = None
+
+        ob_space = env.observation_space
+        X = observ_placeholder if observ_placeholder is not None else observation_placeholder(ob_space,
+                                                                                              batch_size=nbatch)
+        dones = tf.placeholder(tf.float32, shape=[X.shape[0]], name='dones')
+        encoded_x = encode_observation(ob_space, X)
+
+        with tf.variable_scope('current_rnn_memory'):
+            if value_network == 'shared':
+                value_network_ = value_network
+            else:
+                if value_network == 'copy':
+                    value_network_ = policy_network
+                else:
+                    assert callable(value_network)
+                    value_network_ = value_network
+
+            policy_memory_size = policy_network.memory_size if isinstance(policy_network, RNN) else 0
+            value_memory_size = value_network_.memory_size if isinstance(value_network_, RNN) else 0
+            state_size = policy_memory_size + value_memory_size
+
+            if state_size > 0:
+                state_placeholder = tf.placeholder(dtype=tf.float32, shape=(nbatch, state_size),
+                                                   name='states')
+
+                state_map['policy'] = state_placeholder[:, 0:policy_memory_size]
+                state_map['value'] = state_placeholder[:, policy_memory_size:]
+
+        with tf.variable_scope('policy_latent', reuse=tf.AUTO_REUSE):
+            if isinstance(policy_network, RNN):
+                assert policy_memory_size > 0
+                policy_latent, next_policy_state = \
+                    policy_network(encoded_x, dones, state_map['policy'])
+                next_states_list.append(next_policy_state)
+            else:
+                policy_latent = policy_network(encoded_x)
+
+        with tf.variable_scope('value_latent', reuse=tf.AUTO_REUSE):
+            if value_network_ == 'shared':
+                value_latent = policy_latent
+            elif isinstance(value_network_, RNN):
+                assert value_memory_size > 0
+                value_latent, next_value_state = \
+                    value_network_(encoded_x, dones, state_map['value'])
+                next_states_list.append(next_value_state)
+            else:
+                value_latent = value_network_(encoded_x)
+
+        with tf.name_scope("next_rnn_memory"):
+            if state_size > 0:
+                next_states = tf.concat(next_states_list, axis=1)
+                state_info = {'current': state_placeholder,
+                              'next': next_states, }
+            else:
+                state_info = None
+
+        policy = PolicyWithValue(
+            env=env,
+            observations=X,
+            dones=dones,
+            latent=policy_latent,
+            vf_latent=value_latent,
+            states=state_info,
+            sess=sess,
+            estimate_q=estimate_q,
+        )
+        return policy
+
+    return policy_fn

baselines/ppo2/ppo2.py

@@ -1,28 +1,35 @@
 import os
-import time
-import numpy as np
 import os.path as osp
-from baselines import logger
+import time
 from collections import deque
-from baselines.common import explained_variance, set_global_seeds
+
-from baselines.common.policies import build_policy
+import numpy as np
+import tensorflow as tf
+from baselines import logger
+from baselines.common import explained_variance
+from baselines.common import set_global_seeds
+from baselines.common.tf_util import display_var_info
+from baselines.ppo2.policies import build_ppo_policy
+from baselines.ppo2.runner import Runner
+
 try:
     from mpi4py import MPI
 except ImportError:
     MPI = None
-from baselines.ppo2.runner import Runner
 
 
 def constfn(val):
     def f(_):
         return val
+
     return f
 
-def learn(*, network, env, total_timesteps, eval_env = None, seed=None, nsteps=2048, ent_coef=0.0, lr=3e-4,
+
+def learn(*, network, env, total_timesteps, eval_env=None, seed=None, nsteps=128, ent_coef=0.0, lr=3e-4,
           vf_coef=0.5, max_grad_norm=0.5, gamma=0.99, lam=0.95,
           log_interval=10, nminibatches=4, noptepochs=4, cliprange=0.2,
-            save_interval=0, load_path=None, model_fn=None, **network_kwargs):
+          save_interval=10, load_path=None, model_fn=None, **network_kwargs):
-    '''
+    """
     Learn policy using PPO algorithm (https://arxiv.org/abs/1707.06347)
 
     Parameters:
@@ -52,7 +59,7 @@ def learn(*, network, env, total_timesteps, eval_env = None, seed=None, nsteps=2
 
     max_grad_norm: float or None      gradient norm clipping coefficient
 
-    gamma: float                      discounting factor
+    gamma: float                      discounting factor for rewards
 
     lam: float                        advantage estimation discounting factor (lambda in the paper)
 
@@ -72,20 +79,21 @@ def learn(*, network, env, total_timesteps, eval_env = None, seed=None, nsteps=2
 
     **network_kwargs:                 keyword arguments to the policy / network builder. See baselines.common/policies.py/build_policy and arguments to a particular type of network
                                       For instance, 'mlp' network architecture has arguments num_hidden and num_layers.
-
+    """
-
-
-    '''
 
     set_global_seeds(seed)
 
-    if isinstance(lr, float): lr = constfn(lr)
+    if isinstance(lr, float):
-    else: assert callable(lr)
+        lr = constfn(lr)
-    if isinstance(cliprange, float): cliprange = constfn(cliprange)
+    else:
-    else: assert callable(cliprange)
+        assert callable(lr)
+    if isinstance(cliprange, float):
+        cliprange = constfn(cliprange)
+    else:
+        assert callable(cliprange)
     total_timesteps = int(total_timesteps)
 
-    policy = build_policy(env, network, **network_kwargs)
+    policy = build_ppo_policy(env, network, **network_kwargs)
 
     # Get the nb of env
     nenvs = env.num_envs
@@ -104,15 +112,19 @@ def learn(*, network, env, total_timesteps, eval_env = None, seed=None, nsteps=2
         model_fn = Model
 
     model = model_fn(policy=policy, ob_space=ob_space, ac_space=ac_space, nbatch_act=nenvs, nbatch_train=nbatch_train,
-                    nsteps=nsteps, ent_coef=ent_coef, vf_coef=vf_coef,
+                     nsteps=nsteps, ent_coef=ent_coef, vf_coef=vf_coef, max_grad_norm=max_grad_norm)
-                    max_grad_norm=max_grad_norm)
 
     if load_path is not None:
         model.load(load_path)
+
+    allvars = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope=model.name)
+    display_var_info(allvars)
+
     # Instantiate the runner object
-    runner = Runner(env=env, model=model, nsteps=nsteps, gamma=gamma, lam=lam)
+    runner = Runner(env=env, model=model, nsteps=nsteps, gamma=gamma, ob_space=ob_space, lam=lam)
+
     if eval_env is not None:
-        eval_runner = Runner(env = eval_env, model = model, nsteps = nsteps, gamma = gamma, lam= lam)
+        eval_runner = Runner(env=eval_env, model=model, nsteps=nsteps, gamma=gamma, ob_space=ob_space, lam=lam)
 
     epinfobuf = deque(maxlen=100)
     if eval_env is not None:
@@ -120,8 +132,8 @@ def learn(*, network, env, total_timesteps, eval_env = None, seed=None, nsteps=2
 
     # Start total timer
     tfirststart = time.perf_counter()
-
     nupdates = total_timesteps // nbatch
+
     for update in range(1, nupdates + 1):
         assert nbatch % nminibatches == 0
         # Start timer
@@ -131,18 +143,28 @@ def learn(*, network, env, total_timesteps, eval_env = None, seed=None, nsteps=2
         lrnow = lr(frac)
         # Calculate the cliprange
         cliprangenow = cliprange(frac)
-        # Get minibatch
-        obs, returns, masks, actions, values, neglogpacs, states, epinfos = runner.run() #pylint: disable=E0632
-        if eval_env is not None:
-            eval_obs, eval_returns, eval_masks, eval_actions, eval_values, eval_neglogpacs, eval_states, eval_epinfos = eval_runner.run() #pylint: disable=E0632
 
-        epinfobuf.extend(epinfos)
+        # Get minibatch
+        minibatch = runner.run()
+
+        if eval_env is not None:
+            eval_minibatch = eval_runner.run()
+            _eval_obs = eval_minibatch['observations']  # noqa: F841
+            _eval_returns = eval_minibatch['returns']   # noqa: F841
+            _eval_masks = eval_minibatch['masks']       # noqa: F841
+            _eval_actions = eval_minibatch['actions']   # noqa: F841
+            _eval_values = eval_minibatch['values']     # noqa: F841
+            _eval_neglogpacs = eval_minibatch['neglogpacs'] # noqa: F841
+            _eval_states = eval_minibatch['state']      # noqa: F841
+            eval_epinfos = eval_minibatch['epinfos']
+
+        epinfobuf.extend(minibatch.pop('epinfos'))
         if eval_env is not None:
             eval_epinfobuf.extend(eval_epinfos)
 
         # Here what we're going to do is for each minibatch calculate the loss and append it.
         mblossvals = []
-        if states is None: # nonrecurrent version
+
         # Index of each element of batch_size
         # Create the indices array
         inds = np.arange(nbatch)
@@ -153,22 +175,8 @@ def learn(*, network, env, total_timesteps, eval_env = None, seed=None, nsteps=2
             for start in range(0, nbatch, nbatch_train):
                 end = start + nbatch_train
                 mbinds = inds[start:end]
-                    slices = (arr[mbinds] for arr in (obs, returns, masks, actions, values, neglogpacs))
+                slices = {key: minibatch[key][mbinds] for key in minibatch}
-                    mblossvals.append(model.train(lrnow, cliprangenow, *slices))
+                mblossvals.append(model.train(lrnow, cliprangenow, **slices))
-        else: # recurrent version
-            assert nenvs % nminibatches == 0
-            envsperbatch = nenvs // nminibatches
-            envinds = np.arange(nenvs)
-            flatinds = np.arange(nenvs * nsteps).reshape(nenvs, nsteps)
-            for _ in range(noptepochs):
-                np.random.shuffle(envinds)
-                for start in range(0, nenvs, envsperbatch):
-                    end = start + envsperbatch
-                    mbenvinds = envinds[start:end]
-                    mbflatinds = flatinds[mbenvinds].ravel()
-                    slices = (arr[mbflatinds] for arr in (obs, returns, masks, actions, values, neglogpacs))
-                    mbstates = states[mbenvinds]
-                    mblossvals.append(model.train(lrnow, cliprangenow, *slices, mbstates))
 
         # Feedforward --> get losses --> update
         lossvals = np.mean(mblossvals, axis=0)
@@ -179,7 +187,7 @@ def learn(*, network, env, total_timesteps, eval_env = None, seed=None, nsteps=2
         if update % log_interval == 0 or update == 1:
             # Calculates if value function is a good predicator of the returns (ev > 1)
             # or if it's just worse than predicting nothing (ev =< 0)
-            ev = explained_variance(values, returns)
+            ev = explained_variance(minibatch['values'], minibatch['returns'])
             logger.logkv("serial_timesteps", update * nsteps)
             logger.logkv("nupdates", update)
             logger.logkv("total_timesteps", update * nbatch)
@@ -187,6 +195,9 @@ def learn(*, network, env, total_timesteps, eval_env = None, seed=None, nsteps=2
             logger.logkv("explained_variance", float(ev))
             logger.logkv('eprewmean', safemean([epinfo['r'] for epinfo in epinfobuf]))
             logger.logkv('eplenmean', safemean([epinfo['l'] for epinfo in epinfobuf]))
+            logger.logkv('rewards_per_step', safemean(minibatch['rewards']))
+            logger.logkv('advantages_per_step', safemean(minibatch['advs']))
+
             if eval_env is not None:
                 logger.logkv('eval_eprewmean', safemean([epinfo['r'] for epinfo in eval_epinfobuf]))
                 logger.logkv('eval_eplenmean', safemean([epinfo['l'] for epinfo in eval_epinfobuf]))
@@ -195,16 +206,17 @@ def learn(*, network, env, total_timesteps, eval_env = None, seed=None, nsteps=2
                 logger.logkv(lossname, lossval)
             if MPI is None or MPI.COMM_WORLD.Get_rank() == 0:
                 logger.dumpkvs()
-        if save_interval and (update % save_interval == 0 or update == 1) and logger.get_dir() and (MPI is None or MPI.COMM_WORLD.Get_rank() == 0):
+        if save_interval and (update % save_interval == 0 or update == 1) and logger.get_dir() and (
+            MPI is None or MPI.COMM_WORLD.Get_rank() == 0):
             checkdir = osp.join(logger.get_dir(), 'checkpoints')
             os.makedirs(checkdir, exist_ok=True)
             savepath = osp.join(checkdir, '%.5i' % update)
             print('Saving to', savepath)
             model.save(savepath)
+        del minibatch
     return model
+
+
 # Avoid division error when calculate the mean (in our case if epinfo is empty returns np.nan, not return an error)
 def safemean(xs):
     return np.nan if len(xs) == 0 else np.mean(xs)
-
-
-

baselines/ppo2/runner.py

@@ -1,6 +1,8 @@
 import numpy as np
+
 from baselines.common.runners import AbstractEnvRunner
 
+
 class Runner(AbstractEnvRunner):
     """
     We use this object to make a mini batch of experiences
@@ -10,67 +12,118 @@ class Runner(AbstractEnvRunner):
     run():
     - Make a mini batch
     """
-    def __init__(self, *, env, model, nsteps, gamma, lam):
+
+    def __init__(self, *, env, model, nsteps, gamma, ob_space, lam):
         super().__init__(env=env, model=model, nsteps=nsteps)
-        # Lambda used in GAE (General Advantage Estimation)
+
-        self.lam = lam
+        self.lam = lam  # Lambda used in GAE (General Advantage Estimation)
-        # Discount rate
+        self.gamma = gamma  # Discount rate for rewards
-        self.gamma = gamma
+        self.ob_space = ob_space
 
     def run(self):
-        # Here, we init the lists that will contain the mb of experiences
+        minibatch = {
-        mb_obs, mb_rewards, mb_actions, mb_values, mb_dones, mb_neglogpacs = [],[],[],[],[],[]
+            "observations": [],
-        mb_states = self.states
+            "actions": [],
+            "rewards": [],
+            "values": [],
+            "dones": [],
+            "neglogpacs": [],
+        }
+
+        data_type = {
+            "observations": self.obs.dtype,
+            "actions": np.float32,
+            "rewards": np.float32,
+            "values": np.float32,
+            "dones": np.float32,
+            "neglogpacs": np.float32,
+        }
+
+        prev_transition = {'next_states': self.model.initial_state} if self.model.initial_state is not None else {}
         epinfos = []
+
         # For n in range number of steps
         for _ in range(self.nsteps):
-            # Given observations, get action value and neglopacs
+            transitions = {}
-            # We already have self.obs because Runner superclass run self.obs[:] = env.reset() on init
+            transitions['observations'] = self.obs.copy()
-            actions, values, self.states, neglogpacs = self.model.step(self.obs, S=self.states, M=self.dones)
+            transitions['dones'] = self.dones
-            mb_obs.append(self.obs.copy())
+            if 'next_states' in prev_transition:
-            mb_actions.append(actions)
+                transitions['states'] = prev_transition['next_states']
-            mb_values.append(values)
+            transitions.update(self.model.step_with_dict(**transitions))
-            mb_neglogpacs.append(neglogpacs)
-            mb_dones.append(self.dones)
 
             # Take actions in env and look the results
             # Infos contains a ton of useful informations
-            self.obs[:], rewards, self.dones, infos = self.env.step(actions)
+            self.obs, transitions['rewards'], self.dones, infos = self.env.step(transitions['actions'])
+            self.dones = np.array(self.dones, dtype=np.float)
+
             for info in infos:
                 maybeepinfo = info.get('episode')
-                if maybeepinfo: epinfos.append(maybeepinfo)
+                if maybeepinfo:
-            mb_rewards.append(rewards)
+                    epinfos.append(maybeepinfo)
-        #batch of steps to batch of rollouts
-        mb_obs = np.asarray(mb_obs, dtype=self.obs.dtype)
-        mb_rewards = np.asarray(mb_rewards, dtype=np.float32)
-        mb_actions = np.asarray(mb_actions)
-        mb_values = np.asarray(mb_values, dtype=np.float32)
-        mb_neglogpacs = np.asarray(mb_neglogpacs, dtype=np.float32)
-        mb_dones = np.asarray(mb_dones, dtype=np.bool)
-        last_values = self.model.value(self.obs, S=self.states, M=self.dones)
 
-        # discount/bootstrap off value fn
+            for key in transitions:
-        mb_returns = np.zeros_like(mb_rewards)
+                if key not in minibatch:
-        mb_advs = np.zeros_like(mb_rewards)
+                    minibatch[key] = []
-        lastgaelam = 0
+                minibatch[key].append(transitions[key])
+            prev_transition = transitions
+
+        for key in minibatch:
+            dtype = data_type[key] if key in data_type else np.float
+            minibatch[key] = np.array(minibatch[key], dtype=dtype)
+
+        transitions['observations'] = self.obs.copy()
+        transitions['dones'] = self.dones
+        if 'states' in transitions:
+            transitions['states'] = transitions.pop('next_states')
+
+        for key in minibatch:
+            dtype = data_type[key] if key in data_type else np.float
+            minibatch[key] = np.asarray(minibatch[key], dtype=dtype)
+
+        last_values = self.model.step_with_dict(**transitions)['values']
+
+        # Calculate returns and advantages.
+        minibatch['advs'], minibatch['returns'] = \
+            self.advantage_and_returns(values=minibatch['values'],
+                                       rewards=minibatch['rewards'],
+                                       dones=minibatch['dones'],
+                                       last_values=last_values,
+                                       last_dones=self.dones,
+                                       gamma=self.gamma)
+
+        for key in minibatch:
+            minibatch[key] = sf01(minibatch[key])
+
+        minibatch['epinfos'] = epinfos
+        return minibatch
+
+    def advantage_and_returns(self, values, rewards, dones, last_values, last_dones, gamma,
+                              use_non_episodic_rewards=False):
+        """
+        calculate Generalized Advantage Estimation (GAE), https://arxiv.org/abs/1506.02438
+        see also Proximal Policy Optimization Algorithms, https://arxiv.org/abs/1707.06347
+        """
+
+        advantages = np.zeros_like(rewards)
+        lastgaelam = 0  # Lambda used in General Advantage Estimation
         for t in reversed(range(self.nsteps)):
+            if not use_non_episodic_rewards:
                 if t == self.nsteps - 1:
-                nextnonterminal = 1.0 - self.dones
+                    next_non_terminal = 1.0 - last_dones
-                nextvalues = last_values
                 else:
-                nextnonterminal = 1.0 - mb_dones[t+1]
+                    next_non_terminal = 1.0 - dones[t + 1]
-                nextvalues = mb_values[t+1]
+            else:
-            delta = mb_rewards[t] + self.gamma * nextvalues * nextnonterminal - mb_values[t]
+                next_non_terminal = 1.0
-            mb_advs[t] = lastgaelam = delta + self.gamma * self.lam * nextnonterminal * lastgaelam
+            next_value = values[t + 1] if t < self.nsteps - 1 else last_values
-        mb_returns = mb_advs + mb_values
+            delta = rewards[t] + gamma * next_value * next_non_terminal - values[t]
-        return (*map(sf01, (mb_obs, mb_returns, mb_dones, mb_actions, mb_values, mb_neglogpacs)),
+            advantages[t] = lastgaelam = delta + gamma * self.lam * next_non_terminal * lastgaelam
-            mb_states, epinfos)
+        returns = advantages + values
-# obs, returns, masks, actions, values, neglogpacs, states = runner.run()
+        return advantages, returns
+
+
 def sf01(arr):
     """
     swap and then flatten axes 0 and 1
     """
     s = arr.shape
     return arr.swapaxes(0, 1).reshape(s[0] * s[1], *s[2:])
-
-