mirror of
https://github.com/gryf/coach.git
synced 2025-12-18 11:40:18 +01:00
pre-release 0.10.0
This commit is contained in:
Gal Novik
@@ -0,0 +1,114 @@
|
||||
#
|
||||
# Copyright (c) 2017 Intel Corporation
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
#
|
||||
|
||||
from typing import List, Union
|
||||
|
||||
import numpy as np
|
||||
import tensorflow as tf
|
||||
from rl_coach.architectures.tensorflow_components.shared_variables import SharedRunningStats
|
||||
from rl_coach.base_parameters import EmbedderScheme
|
||||
|
||||
from rl_coach.architectures.tensorflow_components.architecture import batchnorm_activation_dropout
|
||||
from rl_coach.core_types import InputEmbedding
|
||||
|
||||
|
||||
class InputEmbedder(object):
|
||||
"""
|
||||
An input embedder is the first part of the network, which takes the input from the state and produces a vector
|
||||
embedding by passing it through a neural network. The embedder will mostly be input type dependent, and there
|
||||
can be multiple embedders in a single network
|
||||
"""
|
||||
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):
|
||||
self.name = name
|
||||
self.input_size = input_size
|
||||
self.activation_function = activation_function
|
||||
self.batchnorm = batchnorm
|
||||
self.dropout = dropout
|
||||
self.dropout_rate = 0
|
||||
self.input = None
|
||||
self.output = None
|
||||
self.scheme = scheme
|
||||
self.return_type = InputEmbedding
|
||||
self.layers = []
|
||||
self.input_rescaling = input_rescaling
|
||||
self.input_offset = input_offset
|
||||
self.input_clipping = input_clipping
|
||||
|
||||
def __call__(self, prev_input_placeholder=None):
|
||||
with tf.variable_scope(self.get_name()):
|
||||
if prev_input_placeholder is None:
|
||||
self.input = tf.placeholder("float", shape=[None] + self.input_size, name=self.get_name())
|
||||
else:
|
||||
self.input = prev_input_placeholder
|
||||
self._build_module()
|
||||
|
||||
return self.input, self.output
|
||||
|
||||
def _build_module(self):
|
||||
# NOTE: for image inputs, we expect the data format to be of type uint8, so to be memory efficient. we chose not
|
||||
# to implement the rescaling as an input filters.observation.observation_filter, as this would have caused the
|
||||
# input to the network to be float, which is 4x more expensive in memory.
|
||||
# thus causing each saved transition in the memory to also be 4x more pricier.
|
||||
|
||||
input_layer = self.input / self.input_rescaling
|
||||
input_layer -= self.input_offset
|
||||
# clip input using te given range
|
||||
if self.input_clipping is not None:
|
||||
input_layer = tf.clip_by_value(input_layer, self.input_clipping[0], self.input_clipping[1])
|
||||
|
||||
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))
|
||||
|
||||
self.output = tf.contrib.layers.flatten(self.layers[-1])
|
||||
|
||||
@property
|
||||
def input_size(self) -> List[int]:
|
||||
return self._input_size
|
||||
|
||||
@input_size.setter
|
||||
def input_size(self, value: Union[int, List[int]]):
|
||||
if isinstance(value, np.ndarray) or isinstance(value, tuple):
|
||||
value = list(value)
|
||||
elif isinstance(value, int):
|
||||
value = [value]
|
||||
if not isinstance(value, list):
|
||||
raise ValueError((
|
||||
'input_size expected to be a list, found {value} which has type {type}'
|
||||
).format(value=value, type=type(value)))
|
||||
self._input_size = value
|
||||
|
||||
@property
|
||||
def schemes(self):
|
||||
raise NotImplementedError("Inheriting embedder must define schemes matching its allowed default "
|
||||
"configurations.")
|
||||
|
||||
def get_name(self):
|
||||
return self.name
|
||||
@@ -0,0 +1,74 @@
|
||||
#
|
||||
# Copyright (c) 2017 Intel Corporation
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
#
|
||||
|
||||
from typing import List
|
||||
|
||||
import tensorflow as tf
|
||||
from rl_coach.architectures.tensorflow_components.architecture import Conv2d
|
||||
from rl_coach.base_parameters import EmbedderScheme
|
||||
|
||||
from rl_coach.architectures.tensorflow_components.embedders.embedder import InputEmbedder
|
||||
from rl_coach.core_types import InputImageEmbedding
|
||||
|
||||
|
||||
class ImageEmbedder(InputEmbedder):
|
||||
"""
|
||||
An input embedder that performs convolutions on the input and then flattens the result.
|
||||
The embedder is intended for image like inputs, where the channels are expected to be the last axis.
|
||||
The embedder also allows custom rescaling of the input prior to the neural network.
|
||||
"""
|
||||
schemes = {
|
||||
EmbedderScheme.Empty:
|
||||
[],
|
||||
|
||||
EmbedderScheme.Shallow:
|
||||
[
|
||||
Conv2d([32, 3, 1])
|
||||
],
|
||||
|
||||
# atari dqn
|
||||
EmbedderScheme.Medium:
|
||||
[
|
||||
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])
|
||||
]
|
||||
}
|
||||
|
||||
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):
|
||||
super().__init__(input_size, activation_function, scheme, batchnorm, dropout, name, input_rescaling,
|
||||
input_offset, input_clipping)
|
||||
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: {}"
|
||||
.format(input_size))
|
||||
|
||||
|
||||
@@ -0,0 +1,64 @@
|
||||
#
|
||||
# Copyright (c) 2017 Intel Corporation
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
#
|
||||
|
||||
from typing import List
|
||||
|
||||
import tensorflow as tf
|
||||
from rl_coach.architectures.tensorflow_components.architecture import Dense
|
||||
from rl_coach.base_parameters import EmbedderScheme
|
||||
|
||||
from rl_coach.architectures.tensorflow_components.embedders.embedder import InputEmbedder
|
||||
from rl_coach.core_types import InputVectorEmbedding
|
||||
|
||||
|
||||
class VectorEmbedder(InputEmbedder):
|
||||
"""
|
||||
An input embedder that is intended for inputs that can be represented as vectors.
|
||||
The embedder flattens the input, applies several dense layers to it and returns the output.
|
||||
"""
|
||||
schemes = {
|
||||
EmbedderScheme.Empty:
|
||||
[],
|
||||
|
||||
EmbedderScheme.Shallow:
|
||||
[
|
||||
Dense([128])
|
||||
],
|
||||
|
||||
# dqn
|
||||
EmbedderScheme.Medium:
|
||||
[
|
||||
Dense([256])
|
||||
],
|
||||
|
||||
# carla
|
||||
EmbedderScheme.Deep: \
|
||||
[
|
||||
Dense([128]),
|
||||
Dense([128]),
|
||||
Dense([128])
|
||||
]
|
||||
}
|
||||
|
||||
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):
|
||||
super().__init__(input_size, activation_function, scheme, batchnorm, dropout, name,
|
||||
input_rescaling, input_offset, input_clipping)
|
||||
|
||||
self.return_type = InputVectorEmbedding
|
||||
if len(self.input_size) != 1 and scheme != EmbedderScheme.Empty:
|
||||
raise ValueError("The input size of a vector embedder must contain only a single dimension")
|
||||
Reference in New Issue
Block a user