microbatch fixes and test (#169)

* microbatch fixes and test

* tiny cleanup

* added assertions to the test

* vpg-related fix

baselines/common/tests/test_serialization.py

@@ -103,9 +103,9 @@ def test_coexistence(learn_fn, network_fn):
     kwargs.update(learn_kwargs[learn_fn])
 
     learn =  partial(learn, env=env, network=network_fn, total_timesteps=0, **kwargs)
-    make_session(make_default=True, graph=tf.Graph());
+    make_session(make_default=True, graph=tf.Graph())
     model1 = learn(seed=1)
-    make_session(make_default=True, graph=tf.Graph());
+    make_session(make_default=True, graph=tf.Graph())
     model2 = learn(seed=2)
 
     model1.step(env.observation_space.sample())

baselines/ppo2/ppo2.py

@@ -131,27 +131,39 @@ class Model(object):
 
             # Normalize the advantages
             advs = (advs - advs.mean()) / (advs.std() + 1e-8)
+
+            if microbatch_size == None or microbatch_size == obs.shape[0]:
                 td_map = {train_model.X:obs, A:actions, ADV:advs, R:returns, LR:lr,
                     CLIPRANGE:cliprange, OLDNEGLOGPAC:neglogpacs, OLDVPRED:values}
                 if states is not None:
                     td_map[train_model.S] = states
                     td_map[train_model.M] = masks
 
-            if microbatch_size == None or microbatch_size == obs.shape[0]:
                 return sess.run(
                     [pg_loss, vf_loss, entropy, approxkl, clipfrac, _train],
                     td_map
                 )[:-1]
             else:
-                sum_grad_v = []
+                assert states is None, "microbatches with recurrent models are not supported yet"
                 pg_losses = []
                 vf_losses = []
                 entropies = []
                 approx_kls = []
                 clipfracs = []
-                for _ in range(nmicrobatches):
+                for microbatch_idx in range(nmicrobatches):
+                    _sli = range(microbatch_idx * microbatch_size, (microbatch_idx+1) * microbatch_size)
+                    td_map = {
+                        train_model.X: obs[_sli],
+                        A:actions[_sli],
+                        ADV:advs[_sli],
+                        R:returns[_sli],
+                        CLIPRANGE:cliprange,
+                        OLDNEGLOGPAC:neglogpacs[_sli],
+                        OLDVPRED:values[_sli]
+                    }
+
                     grad_v, pg_loss_v, vf_loss_v, entropy_v, approx_kl_v, clipfrac_v  = sess.run([grads, pg_loss, vf_loss, entropy, approxkl, clipfrac], td_map)
-                    if len(sum_grad_v) == 0:
+                    if microbatch_idx == 0:
                         sum_grad_v = grad_v
                     else:
                         for i, g in enumerate(grad_v):

baselines/ppo2/test_microbatches.py

@@ -0,0 +1,28 @@
+import gym
+import tensorflow as tf
+import numpy as np
+from baselines.common.vec_env.dummy_vec_env import DummyVecEnv
+from baselines.common.tf_util import make_session
+from baselines.ppo2.ppo2 import learn
+
+def test_microbatches():
+    def env_fn():
+        env = gym.make('CartPole-v0')
+        env.seed(0)
+        return env
+
+    env_ref = DummyVecEnv([env_fn])
+    sess_ref = make_session(make_default=True, graph=tf.Graph())
+    learn(env=env_ref, network='mlp', nsteps=32, total_timesteps=32, seed=0)
+    vars_ref = {v.name: sess_ref.run(v) for v in tf.trainable_variables()}
+
+    env_test = DummyVecEnv([env_fn])
+    sess_test = make_session(make_default=True, graph=tf.Graph())
+    learn(env=env_test, network='mlp', nsteps=32, total_timesteps=32, seed=0, microbatch_size=4)
+    vars_test = {v.name: sess_test.run(v) for v in tf.trainable_variables()}
+
+    for v in vars_ref:
+        np.testing.assert_allclose(vars_ref[v], vars_test[v], atol=1e-3)
+
+if __name__ == '__main__':
+    test_microbatches()