gryf/coach
1
0
Fork 0
mirror of https://github.com/gryf/coach.git synced 2026-03-03 15:25:49 +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

59
rl_coach/architectures/tensorflow_components/embedders/embedder.py
View File

@@ -15,20 +15,23 @@
#
from typing import List, Union
import copy
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, \
BatchnormActivationDropout
from rl_coach.base_parameters import EmbedderScheme, NetworkComponentParameters
from rl_coach.core_types import InputEmbedding
from rl_coach.utils import force_list
class InputEmbedderParameters(NetworkComponentParameters):
def __init__(self, activation_function: str='relu', scheme: Union[List, EmbedderScheme]=EmbedderScheme.Medium,
batchnorm: bool=False, dropout=False, name: str='embedder', input_rescaling=None, input_offset=None,
input_clipping=None, dense_layer=Dense):
input_clipping=None, dense_layer=Dense, is_training=False):
super().__init__(dense_layer=dense_layer)
self.activation_function = activation_function
self.scheme = scheme
@@ -44,6 +47,7 @@ class InputEmbedderParameters(NetworkComponentParameters):
self.input_offset = input_offset
self.input_clipping = input_clipping
self.name = name
self.is_training = is_training
@property
def path(self):
@@ -61,7 +65,8 @@ class InputEmbedder(object):
"""
def __init__(self, input_size: List[int], activation_function=tf.nn.relu,
scheme: EmbedderScheme=None, batchnorm: bool=False, dropout: bool=False,
name: str= "embedder", input_rescaling=1.0, input_offset=0.0, input_clipping=None, dense_layer=Dense):
name: str= "embedder", input_rescaling=1.0, input_offset=0.0, input_clipping=None, dense_layer=Dense,
is_training=False):
self.name = name
self.input_size = input_size
self.activation_function = activation_function
@@ -72,11 +77,29 @@ class InputEmbedder(object):
self.output = None
self.scheme = scheme
self.return_type = InputEmbedding
self.layers_params = []
self.layers = []
self.input_rescaling = input_rescaling
self.input_offset = input_offset
self.input_clipping = input_clipping
self.dense_layer = dense_layer
self.is_training = is_training
# layers order is conv -> batchnorm -> activation -> dropout
if isinstance(self.scheme, EmbedderScheme):
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, prev_input_placeholder=None):
with tf.variable_scope(self.get_name()):
@@ -102,19 +125,11 @@ class InputEmbedder(object):
self.layers.append(input_layer)
# layers order is conv -> batchnorm -> activation -> dropout
if isinstance(self.scheme, EmbedderScheme):
layers_params = self.schemes[self.scheme]
else:
layers_params = self.scheme
for idx, layer_params in enumerate(layers_params):
self.layers.append(
layer_params(input_layer=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(input_layer=self.layers[-1], name='{}_{}'.format(layer_params.__class__.__name__, idx),
is_training=self.is_training)
))
self.output = tf.contrib.layers.flatten(self.layers[-1])
@@ -140,4 +155,14 @@ class InputEmbedder(object):
"configurations.")
def get_name(self):
return self.name
return self.name
def __str__(self):
result = []
if self.input_rescaling != 1.0 or self.input_offset != 0.0:
result.append('Input Normalization (scale = {}, offset = {})'.format(self.input_rescaling, self.input_offset))
result.extend([str(l) for l in self.layers_params])
if self.layers_params:
return '\n'.join(result)
else:
return 'No layers'

30
rl_coach/architectures/tensorflow_components/embedders/image_embedder.py
View File

@@ -18,7 +18,7 @@ from typing import List
import tensorflow as tf
from rl_coach.architectures.tensorflow_components.architecture import Conv2d, Dense
from rl_coach.architectures.tensorflow_components.layers import Conv2d, Dense
from rl_coach.architectures.tensorflow_components.embedders.embedder import InputEmbedder
from rl_coach.base_parameters import EmbedderScheme
from rl_coach.core_types import InputImageEmbedding
@@ -34,9 +34,9 @@ class ImageEmbedder(InputEmbedder):
def __init__(self, input_size: List[int], activation_function=tf.nn.relu,
scheme: EmbedderScheme=EmbedderScheme.Medium, batchnorm: bool=False, dropout: bool=False,
name: str= "embedder", input_rescaling: float=255.0, input_offset: float=0.0, input_clipping=None,
dense_layer=Dense):
dense_layer=Dense, is_training=False):
super().__init__(input_size, activation_function, scheme, batchnorm, dropout, name, input_rescaling,
input_offset, input_clipping, dense_layer=dense_layer)
input_offset, input_clipping, dense_layer=dense_layer, is_training=is_training)
self.return_type = InputImageEmbedding
if len(input_size) != 3 and scheme != EmbedderScheme.Empty:
raise ValueError("Image embedders expect the input size to have 3 dimensions. The given size is: {}"
@@ -50,28 +50,28 @@ class ImageEmbedder(InputEmbedder):
EmbedderScheme.Shallow:
[
Conv2d([32, 3, 1])
Conv2d(32, 3, 1)
],
# atari dqn
EmbedderScheme.Medium:
[
Conv2d([32, 8, 4]),
Conv2d([64, 4, 2]),
Conv2d([64, 3, 1])
Conv2d(32, 8, 4),
Conv2d(64, 4, 2),
Conv2d(64, 3, 1)
],
# carla
EmbedderScheme.Deep: \
[
Conv2d([32, 5, 2]),
Conv2d([32, 3, 1]),
Conv2d([64, 3, 2]),
Conv2d([64, 3, 1]),
Conv2d([128, 3, 2]),
Conv2d([128, 3, 1]),
Conv2d([256, 3, 2]),
Conv2d([256, 3, 1])
Conv2d(32, 5, 2),
Conv2d(32, 3, 1),
Conv2d(64, 3, 2),
Conv2d(64, 3, 1),
Conv2d(128, 3, 2),
Conv2d(128, 3, 1),
Conv2d(256, 3, 2),
Conv2d(256, 3, 1)
]
}

17
rl_coach/architectures/tensorflow_components/embedders/vector_embedder.py
View File

@@ -18,7 +18,7 @@ from typing import List
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.embedders.embedder import InputEmbedder
from rl_coach.base_parameters import EmbedderScheme
from rl_coach.core_types import InputVectorEmbedding
@@ -33,9 +33,10 @@ class VectorEmbedder(InputEmbedder):
def __init__(self, input_size: List[int], activation_function=tf.nn.relu,
scheme: EmbedderScheme=EmbedderScheme.Medium, batchnorm: bool=False, dropout: bool=False,
name: str= "embedder", input_rescaling: float=1.0, input_offset:float=0.0, input_clipping=None,
dense_layer=Dense):
dense_layer=Dense, is_training=False):
super().__init__(input_size, activation_function, scheme, batchnorm, dropout, name,
input_rescaling, input_offset, input_clipping, dense_layer=dense_layer)
input_rescaling, input_offset, input_clipping, dense_layer=dense_layer,
is_training=is_training)
self.return_type = InputVectorEmbedding
if len(self.input_size) != 1 and scheme != EmbedderScheme.Empty:
@@ -49,20 +50,20 @@ class VectorEmbedder(InputEmbedder):
EmbedderScheme.Shallow:
[
self.dense_layer([128])
self.dense_layer(128)
],
# dqn
EmbedderScheme.Medium:
[
self.dense_layer([256])
self.dense_layer(256)
],
# carla
EmbedderScheme.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)
]
}