gryf/coach
1
0
Fork 0
mirror of https://github.com/gryf/coach.git synced 2025-12-18 11:40:18 +01:00
Code Issues Projects Releases Wiki Activity

imporved API for getting / setting variables within the graph

Browse Source
This commit is contained in:
Itai Caspi
2017-10-25 16:07:58 +03:00
parent e33b0e8534
commit 1918f16079
5 changed files with 45 additions and 17 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
architectures/tensorflow_components/architecture.py
View File

@@ -288,3 +288,21 @@ class TensorFlowArchitecture(Architecture):
"""
summary_writer = tf.summary.FileWriter(summary_dir)
summary_writer.add_graph(self.sess.graph)
def get_variable_value(self, variable):
"""
Get the value of a variable from the graph
:param variable: the variable
:return: the value of the variable
"""
return self.sess.run(variable)
def set_variable_value(self, assign_op, value, placeholder=None):
"""
Updates the value of a variable.
This requires having an assign operation for the variable, and a placeholder which will provide the value
:param assign_op: an assign operation for the variable
:param value: a value to set the variable to
:param placeholder: a placeholder to hold the given value for injecting it into the variable
"""
self.sess.run(assign_op, feed_dict={placeholder: value})

6
architectures/tensorflow_components/heads.py
View File

@@ -351,15 +351,19 @@ class PPOHead(Head):
self.num_actions = tuning_parameters.env_instance.action_space_size
self.discrete_controls = tuning_parameters.env_instance.discrete_controls
self.output_scale = np.max(tuning_parameters.env_instance.action_space_abs_range)
# kl coefficient and its corresponding assignment operation and placeholder
self.kl_coefficient = tf.Variable(tuning_parameters.agent.initial_kl_coefficient,
trainable=False, name='kl_coefficient')
self.kl_coefficient_ph = tf.placeholder('float', name='kl_coefficient_ph')
self.assign_kl_coefficient = tf.assign(self.kl_coefficient, self.kl_coefficient_ph)
self.kl_cutoff = 2*tuning_parameters.agent.target_kl_divergence
self.high_kl_penalty_coefficient = tuning_parameters.agent.high_kl_penalty_coefficient
self.clip_likelihood_ratio_using_epsilon = tuning_parameters.agent.clip_likelihood_ratio_using_epsilon
self.use_kl_regularization = tuning_parameters.agent.use_kl_regularization
self.beta = tuning_parameters.agent.beta_entropy
def _build_module(self, input_layer):
eps = 1e-15