gryf/coach
1
0
Fork 0
mirror of https://github.com/gryf/coach.git synced 2025-12-18 03:30:19 +01:00
Code Issues Projects Releases Wiki Activity

network_imporvements branch merge

Browse Source
This commit is contained in:
Shadi Endrawis
2018-10-02 13:41:46 +03:00
parent 72ea933384
commit 51726a5b80
110 changed files with 1639 additions and 1161 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
rl_coach/architectures/tensorflow_components/heads/ppo_head.py
View File

@@ -17,7 +17,7 @@
import numpy as np
import tensorflow as tf
from rl_coach.architectures.tensorflow_components.architecture import Dense
from rl_coach.architectures.tensorflow_components.layers import Dense
from rl_coach.architectures.tensorflow_components.heads.head import Head, HeadParameters, normalized_columns_initializer
from rl_coach.base_parameters import AgentParameters
from rl_coach.core_types import ActionProbabilities
@@ -27,9 +27,13 @@ from rl_coach.utils import eps
class PPOHeadParameters(HeadParameters):
def __init__(self, activation_function: str ='tanh', name: str='ppo_head_params', dense_layer=Dense):
def __init__(self, activation_function: str ='tanh', name: str='ppo_head_params',
num_output_head_copies: int = 1, rescale_gradient_from_head_by_factor: float = 1.0,
loss_weight: float = 1.0, dense_layer=Dense):
super().__init__(parameterized_class=PPOHead, activation_function=activation_function, name=name,
dense_layer=dense_layer)
dense_layer=dense_layer, num_output_head_copies=num_output_head_copies,
rescale_gradient_from_head_by_factor=rescale_gradient_from_head_by_factor,
loss_weight=loss_weight)
class PPOHead(Head):
@@ -146,3 +150,15 @@ class PPOHead(Head):
self.old_policy_distribution = tf.contrib.distributions.MultivariateNormalDiag(self.old_policy_mean, self.old_policy_std + eps)
self.output = [self.policy_mean, self.policy_std]
def __str__(self):
action_head_mean_result = []
if isinstance(self.spaces.action, DiscreteActionSpace):
# create a discrete action network (softmax probabilities output)
action_head_mean_result.append("Dense (num outputs = {})".format(len(self.spaces.action.actions)))
action_head_mean_result.append("Softmax")
elif isinstance(self.spaces.action, BoxActionSpace):
# create a continuous action network (bounded mean and stdev outputs)
action_head_mean_result.append("Dense (num outputs = {})".format(self.spaces.action.shape))
return '\n'.join(action_head_mean_result)