gryf/coach
1
0
Fork 0
mirror of https://github.com/gryf/coach.git synced 2025-12-17 19:20:19 +01:00
Code Issues Projects Releases Wiki Activity

fix bug in ddpg

Browse Source
This commit is contained in:
Zach Dwiel
2018-02-16 20:18:03 -05:00
parent 8248caf35e
commit 5cf10e5f52
2 changed files with 2 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
agents/ddpg_agent.py
View File

@@ -54,7 +54,7 @@ class DDPGAgent(ActorCriticAgent):
actions_mean = self.actor_network.online_network.predict(current_states)
critic_online_network = self.critic_network.online_network
# TODO: convert into call to predict, current method ignores lstm middleware for example
action_gradients = self.critic_network.sess.run(critic_online_network.gradients_wrt_inputs[1],
action_gradients = self.critic_network.sess.run(critic_online_network.gradients_wrt_inputs['action'],
feed_dict=critic_online_network._feed_dict({
**current_states,
'action': actions_mean,

3
architectures/tensorflow_components/architecture.py
View File

@@ -108,8 +108,7 @@ class TensorFlowArchitecture(Architecture):
# gradients of the outputs w.r.t. the inputs
# at the moment, this is only used by ddpg
if len(self.outputs) == 1:
# TODO: convert gradients_with_respect_to_inputs into dictionary?
self.gradients_wrt_inputs = [tf.gradients(self.outputs[0], input_ph) for input_ph in self.inputs.values()]
self.gradients_wrt_inputs = {name: tf.gradients(self.outputs[0], input_ph) for name, input_ph in self.inputs.items()}
self.gradients_weights_ph = tf.placeholder('float32', self.outputs[0].shape, 'output_gradient_weights')
self.weighted_gradients = tf.gradients(self.outputs[0], self.trainable_weights, self.gradients_weights_ph)