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

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

39
rl_coach/architectures/tensorflow_components/middlewares/fc_middleware.py
View File

@@ -17,46 +17,41 @@ from typing import Union, List
import tensorflow as tf
from rl_coach.architectures.tensorflow_components.architecture import batchnorm_activation_dropout, Dense
from rl_coach.architectures.tensorflow_components.layers import batchnorm_activation_dropout, Dense
from rl_coach.architectures.tensorflow_components.middlewares.middleware import Middleware, MiddlewareParameters
from rl_coach.base_parameters import MiddlewareScheme
from rl_coach.core_types import Middleware_FC_Embedding
from rl_coach.utils import force_list
class FCMiddlewareParameters(MiddlewareParameters):
def __init__(self, activation_function='relu',
scheme: Union[List, MiddlewareScheme] = MiddlewareScheme.Medium,
batchnorm: bool = False, dropout: bool = False,
name="middleware_fc_embedder", dense_layer=Dense):
name="middleware_fc_embedder", dense_layer=Dense, is_training=False):
super().__init__(parameterized_class=FCMiddleware, activation_function=activation_function,
scheme=scheme, batchnorm=batchnorm, dropout=dropout, name=name, dense_layer=dense_layer)
scheme=scheme, batchnorm=batchnorm, dropout=dropout, name=name, dense_layer=dense_layer,
is_training=is_training)
class FCMiddleware(Middleware):
def __init__(self, activation_function=tf.nn.relu,
scheme: MiddlewareScheme = MiddlewareScheme.Medium,
batchnorm: bool = False, dropout: bool = False,
name="middleware_fc_embedder", dense_layer=Dense):
name="middleware_fc_embedder", dense_layer=Dense, is_training=False):
super().__init__(activation_function=activation_function, batchnorm=batchnorm,
dropout=dropout, scheme=scheme, name=name, dense_layer=dense_layer)
dropout=dropout, scheme=scheme, name=name, dense_layer=dense_layer, is_training=is_training)
self.return_type = Middleware_FC_Embedding
self.layers = []
def _build_module(self):
self.layers.append(self.input)
if isinstance(self.scheme, MiddlewareScheme):
layers_params = self.schemes[self.scheme]
else:
layers_params = self.scheme
for idx, layer_params in enumerate(layers_params):
self.layers.append(
layer_params(self.layers[-1], name='{}_{}'.format(layer_params.__class__.__name__, idx))
)
self.layers.extend(batchnorm_activation_dropout(self.layers[-1], self.batchnorm,
self.activation_function, self.dropout,
self.dropout_rate, idx))
for idx, layer_params in enumerate(self.layers_params):
self.layers.extend(force_list(
layer_params(self.layers[-1], name='{}_{}'.format(layer_params.__class__.__name__, idx),
is_training=self.is_training)
))
self.output = self.layers[-1]
@@ -69,20 +64,20 @@ class FCMiddleware(Middleware):
# ppo
MiddlewareScheme.Shallow:
[
self.dense_layer([64])
self.dense_layer(64)
],
# dqn
MiddlewareScheme.Medium:
[
self.dense_layer([512])
self.dense_layer(512)
],
MiddlewareScheme.Deep: \
[
self.dense_layer([128]),
self.dense_layer([128]),
self.dense_layer([128])
self.dense_layer(128),
self.dense_layer(128),
self.dense_layer(128)
]
}