stop gradient in policy/distributional

examples/ppo_multivariate_normal.py

@@ -0,0 +1,83 @@
+import tensorflow as tf
+import gym
+import numpy as np
+import time
+
+import tensorflow_probability as tfp
+tfd = tfp.distributions  # TODO: use zhusuan.distributions
+
+import tianshou as ts
+
+
+if __name__ == '__main__':
+    env = gym.make('BipedalWalker-v2')
+    observation_dim = env.observation_space.shape
+    action_dim = env.action_space.shape[0]
+
+    clip_param = 0.2
+    num_batches = 10
+    batch_size = 512
+
+    seed = 0
+    np.random.seed(seed)
+    tf.set_random_seed(seed)
+
+    ### 1. build network with pure tf
+    observation_ph = tf.placeholder(tf.float32, shape=(None,) + observation_dim)
+
+    def my_policy():
+        net = tf.layers.dense(observation_ph, 32, activation=tf.nn.tanh)
+        net = tf.layers.dense(net, 32, activation=tf.nn.tanh)
+
+        action_logits = tf.layers.dense(net, action_dim, activation=None)
+        action_dist = tfd.MultivariateNormalDiag(loc=action_logits, scale_diag=[0.2] * action_dim)
+
+        return action_dist, None
+
+    ### 2. build policy, loss, optimizer
+    pi = ts.policy.Distributional(my_policy, observation_placeholder=observation_ph, has_old_net=True)
+
+    ppo_loss_clip = ts.losses.ppo_clip(pi, clip_param)
+
+    total_loss = ppo_loss_clip
+    optimizer = tf.train.AdamOptimizer(1e-4)
+    train_op = optimizer.minimize(total_loss, var_list=list(pi.trainable_variables))
+
+    ### 3. define data collection
+    data_buffer = ts.data.BatchSet()
+
+    data_collector = ts.data.DataCollector(
+        env=env,
+        policy=pi,
+        data_buffer=data_buffer,
+        process_functions=[ts.data.advantage_estimation.full_return],
+        managed_networks=[pi],
+    )
+
+    ### 4. start training
+    config = tf.ConfigProto()
+    config.gpu_options.allow_growth = True
+    with tf.Session(config=config) as sess:
+        sess.run(tf.global_variables_initializer())
+
+        # assign actor to pi_old
+        pi.sync_weights()
+
+        start_time = time.time()
+        for i in range(1000):
+            # collect data
+            data_collector.collect(num_episodes=50)
+
+            # print current return
+            print('Epoch {}:'.format(i))
+            data_buffer.statistics()
+
+            # update network
+            for _ in range(num_batches):
+                feed_dict = data_collector.next_batch(batch_size)
+                sess.run(train_op, feed_dict=feed_dict)
+
+            # assigning pi_old to be current pi
+            pi.sync_weights()
+
+            print('Elapsed time: {:.1f} min'.format((time.time() - start_time) / 60))

setup.py

@@ -81,7 +81,8 @@ setup(
     # your project is installed. For an analysis of "install_requires" vs pip's
     # requirements files see:
     # https://packaging.python.org/en/latest/requirements.html
-    install_requires=['numpy>=1.14.0'],
+    install_requires=['numpy>=1.14.0',
+                      'tensorflow-probability'],
 
     # List additional groups of dependencies here (e.g. development
     # dependencies). You can install these using the following syntax,

tianshou/core/policy/distributional.py

@@ -36,7 +36,7 @@ class Distributional(PolicyBase):
             self.action = action_dist.sample()
 
         weights = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES)
-        self.network_weights = identify_dependent_variables(self.action_dist.log_prob(self.action), weights)
+        self.network_weights = identify_dependent_variables(self.action_dist.log_prob(tf.stop_gradient(self.action)), weights)
         self._trainable_variables = [var for var in self.network_weights
                                     if var in tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES)]
         # deal with target network
@@ -52,7 +52,7 @@ class Distributional(PolicyBase):
             # re-filter to rule out some edge cases
             old_weights = [var for var in old_weights if var.name[:len(net_old_scope)] == net_old_scope]
 
-            self.network_old_weights = identify_dependent_variables(self.action_dist_old.log_prob(self.action_old), old_weights)
+            self.network_old_weights = identify_dependent_variables(self.action_dist_old.log_prob(tf.stop_gradient(self.action_old)), old_weights)
             assert len(self.network_weights) == len(self.network_old_weights)
 
             self.sync_weights_ops = [tf.assign(variable_old, variable)