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 keep_dims -> keepdims

Browse Source
This commit is contained in:
Zach Dwiel
2018-02-16 13:30:31 -05:00
parent 39a28aba95
commit ee6e0bdc3b
3 changed files with 10 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
agents/nec_agent.py
View File

@@ -16,7 +16,8 @@
import numpy as np import numpy as np
from agents.value_optimization_agent import * from agents.value_optimization_agent import ValueOptimizationAgent
from logger import screen
# Neural Episodic Control - https://arxiv.org/pdf/1703.01988.pdf # Neural Episodic Control - https://arxiv.org/pdf/1703.01988.pdf

6
agents/ppo_agent.py
View File

@@ -112,8 +112,12 @@ class PPOAgent(ActorCriticAgent):
current_values = self.critic_network.online_network.predict(current_states_batch) current_values = self.critic_network.online_network.predict(current_states_batch)
targets = current_values * (1 - mix_fraction) + total_return_batch * mix_fraction targets = current_values * (1 - mix_fraction) + total_return_batch * mix_fraction
inputs = copy.copy(current_states_batch)
for input_index, input in enumerate(old_policy_values):
inputs['output_0_{}'.format(input_index)] = input
value_loss = self.critic_network.online_network.\ value_loss = self.critic_network.online_network.\
accumulate_gradients([current_states_batch] + old_policy_values, targets) accumulate_gradients(inputs, targets)
self.critic_network.apply_gradients_to_online_network() self.critic_network.apply_gradients_to_online_network()
if self.tp.distributed: if self.tp.distributed:
self.critic_network.apply_gradients_to_global_network() self.critic_network.apply_gradients_to_global_network()

6
architectures/tensorflow_components/heads.py
View File

@@ -23,7 +23,7 @@ from utils import force_list
def normalized_columns_initializer(std=1.0): def normalized_columns_initializer(std=1.0):
def _initializer(shape, dtype=None, partition_info=None): def _initializer(shape, dtype=None, partition_info=None):
out = np.random.randn(*shape).astype(np.float32) out = np.random.randn(*shape).astype(np.float32)
out *= std / np.sqrt(np.square(out).sum(axis=0, keepdims=True)) out *= std / np.sqrt(np.square(out).sum(axis=0, keep_dims=True))
return tf.constant(out) return tf.constant(out)
return _initializer return _initializer
@@ -250,7 +250,7 @@ class MeasurementsPredictionHead(Head):
name='output') name='output')
action_stream = tf.reshape(action_stream, action_stream = tf.reshape(action_stream,
(tf.shape(action_stream)[0], self.num_actions, self.multi_step_measurements_size)) (tf.shape(action_stream)[0], self.num_actions, self.multi_step_measurements_size))
action_stream = action_stream - tf.reduce_mean(action_stream, reduction_indices=1, keepdims=True) action_stream = action_stream - tf.reduce_mean(action_stream, reduction_indices=1, keep_dims=True)
# merge to future measurements predictions # merge to future measurements predictions
self.output = tf.add(expectation_stream, action_stream, name='output') self.output = tf.add(expectation_stream, action_stream, name='output')
@@ -302,7 +302,7 @@ class DNDQHead(Head):
square_diff = tf.square(dnd_embeddings - tf.expand_dims(input_layer, 1)) square_diff = tf.square(dnd_embeddings - tf.expand_dims(input_layer, 1))
distances = tf.reduce_sum(square_diff, axis=2) + [self.l2_norm_added_delta] distances = tf.reduce_sum(square_diff, axis=2) + [self.l2_norm_added_delta]
weights = 1.0 / distances weights = 1.0 / distances
normalised_weights = weights / tf.reduce_sum(weights, axis=1, keepdims=True) normalised_weights = weights / tf.reduce_sum(weights, axis=1, keep_dims=True)
return tf.reduce_sum(dnd_values * normalised_weights, axis=1) return tf.reduce_sum(dnd_values * normalised_weights, axis=1)