gryf/coach
1
0
Fork 0
mirror of https://github.com/gryf/coach.git synced 2026-03-16 06:33:36 +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)
]
}

41
rl_coach/architectures/tensorflow_components/middlewares/lstm_middleware.py
View File

@@ -18,19 +18,21 @@
import numpy as np
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_LSTM_Embedding
from rl_coach.utils import force_list
class LSTMMiddlewareParameters(MiddlewareParameters):
def __init__(self, activation_function='relu', number_of_lstm_cells=256,
scheme: MiddlewareScheme = MiddlewareScheme.Medium,
batchnorm: bool = False, dropout: bool = False,
name="middleware_lstm_embedder", dense_layer=Dense):
name="middleware_lstm_embedder", dense_layer=Dense, is_training=False):
super().__init__(parameterized_class=LSTMMiddleware, 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)
self.number_of_lstm_cells = number_of_lstm_cells
@@ -38,9 +40,9 @@ class LSTMMiddleware(Middleware):
def __init__(self, activation_function=tf.nn.relu, number_of_lstm_cells: int=256,
scheme: MiddlewareScheme = MiddlewareScheme.Medium,
batchnorm: bool = False, dropout: bool = False,
name="middleware_lstm_embedder", dense_layer=Dense):
name="middleware_lstm_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_LSTM_Embedding
self.number_of_lstm_cells = number_of_lstm_cells
self.layers = []
@@ -57,19 +59,12 @@ class LSTMMiddleware(Middleware):
self.layers.append(self.input)
# optionally insert some dense layers before the LSTM
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(
tf.layers.dense(self.layers[-1], layer_params[0], name='fc{}'.format(idx))
)
self.layers.extend(batchnorm_activation_dropout(self.layers[-1], self.batchnorm,
self.activation_function, self.dropout,
self.dropout_rate, idx))
# optionally insert some layers before the LSTM
for idx, layer_params in enumerate(self.layers_params):
self.layers.extend(force_list(
layer_params(self.layers[-1], name='fc{}'.format(idx),
is_training=self.is_training)
))
# add the LSTM layer
lstm_cell = tf.nn.rnn_cell.BasicLSTMCell(self.number_of_lstm_cells, state_is_tuple=True)
@@ -97,20 +92,20 @@ class LSTMMiddleware(Middleware):
# ppo
MiddlewareScheme.Shallow:
[
[64]
self.dense_layer(64)
],
# dqn
MiddlewareScheme.Medium:
[
[512]
self.dense_layer(512)
],
MiddlewareScheme.Deep: \
[
[128],
[128],
[128]
self.dense_layer(128),
self.dense_layer(128),
self.dense_layer(128)
]
}

35
rl_coach/architectures/tensorflow_components/middlewares/middleware.py
View File

@@ -13,25 +13,27 @@
# See the License for the specific language governing permissions and
# limitations under the License.
#
import copy
from typing import Type, Union, List
import tensorflow as tf
from rl_coach.architectures.tensorflow_components.architecture import Dense
from rl_coach.base_parameters import MiddlewareScheme, Parameters, NetworkComponentParameters
from rl_coach.architectures.tensorflow_components.layers import Dense, BatchnormActivationDropout
from rl_coach.base_parameters import MiddlewareScheme, NetworkComponentParameters
from rl_coach.core_types import MiddlewareEmbedding
class MiddlewareParameters(NetworkComponentParameters):
def __init__(self, parameterized_class: Type['Middleware'],
activation_function: str='relu', scheme: Union[List, MiddlewareScheme]=MiddlewareScheme.Medium,
batchnorm: bool=False, dropout: bool=False, name='middleware', dense_layer=Dense):
batchnorm: bool=False, dropout: bool=False, name='middleware', dense_layer=Dense, is_training=False):
super().__init__(dense_layer=dense_layer)
self.activation_function = activation_function
self.scheme = scheme
self.batchnorm = batchnorm
self.dropout = dropout
self.name = name
self.is_training = is_training
self.parameterized_class_name = parameterized_class.__name__
@@ -43,7 +45,8 @@ class Middleware(object):
"""
def __init__(self, activation_function=tf.nn.relu,
scheme: MiddlewareScheme = MiddlewareScheme.Medium,
batchnorm: bool = False, dropout: bool = False, name="middleware_embedder", dense_layer=Dense):
batchnorm: bool = False, dropout: bool = False, name="middleware_embedder", dense_layer=Dense,
is_training=False):
self.name = name
self.input = None
self.output = None
@@ -54,6 +57,23 @@ class Middleware(object):
self.scheme = scheme
self.return_type = MiddlewareEmbedding
self.dense_layer = dense_layer
self.is_training = is_training
# layers order is conv -> batchnorm -> activation -> dropout
if isinstance(self.scheme, MiddlewareScheme):
self.layers_params = copy.copy(self.schemes[self.scheme])
else:
self.layers_params = copy.copy(self.scheme)
# we allow adding batchnorm, dropout or activation functions after each layer.
# The motivation is to simplify the transition between a network with batchnorm and a network without
# batchnorm to a single flag (the same applies to activation function and dropout)
if self.batchnorm or self.activation_function or self.dropout:
for layer_idx in reversed(range(len(self.layers_params))):
self.layers_params.insert(layer_idx+1,
BatchnormActivationDropout(batchnorm=self.batchnorm,
activation_function=self.activation_function,
dropout_rate=self.dropout_rate))
def __call__(self, input_layer):
with tf.variable_scope(self.get_name()):
@@ -72,3 +92,10 @@ class Middleware(object):
def schemes(self):
raise NotImplementedError("Inheriting middleware must define schemes matching its allowed default "
"configurations.")
def __str__(self):
result = [str(l) for l in self.layers_params]
if self.layers_params:
return '\n'.join(result)
else:
return 'No layers'