Skip to content

GitLab

"torchvision/csrc/vscode:/vscode.git/clone" did not exist on "6db54fb7eb507d26a5670611398f689c463f30ad"
Commit f5fc733a authored by Byzantine's avatar Byzantine
Browse files

Removing research/community models

parent 09bc9f54
Hide whitespace changes
Inline Side-by-side
Showing with 0 additions and 2159 deletions
research/compression/entropy_coder/core/code_loader.py deleted 100644 → 0
View file @ 09bc9f54
# Copyright 2017 The TensorFlow Authors All Rights Reserved.
#
# 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.
# ==============================================================================
"""Load binary codes stored as tf.Example in a TFRecord table."""
import tensorflow as tf
def ReadFirstCode(dataset):
"""Read the first example from a binary code RecordIO table."""
for record in tf.python_io.tf_record_iterator(dataset):
tf_example = tf.train.Example()
tf_example.ParseFromString(record)
break
return tf_example
def LoadBinaryCode(input_config, batch_size):
"""Load a batch of binary codes from a tf.Example dataset.
Args:
input_config: An InputConfig proto containing the input configuration.
batch_size: Output batch size of examples.
Returns:
A batched tensor of binary codes.
"""
data = input_config.data
# TODO: Possibly use multiple files (instead of just one).
file_list = [data]
filename_queue = tf.train.string_input_producer(file_list,
capacity=4)
reader = tf.TFRecordReader()
_, values = reader.read(filename_queue)
serialized_example = tf.reshape(values, shape=[1])
serialized_features = {
'code_shape': tf.FixedLenFeature([3],
dtype=tf.int64),
'code': tf.VarLenFeature(tf.float32),
}
example = tf.parse_example(serialized_example, serialized_features)
# 3D shape: height x width x binary_code_depth
z = example['code_shape']
code_shape = tf.reshape(tf.cast(z, tf.int32), [3])
# Un-flatten the binary codes.
code = tf.reshape(tf.sparse_tensor_to_dense(example['code']), code_shape)
queue_size = 10
queue = tf.PaddingFIFOQueue(
queue_size + 3 * batch_size,
dtypes=[code.dtype],
shapes=[[None, None, None]])
enqueue_op = queue.enqueue([code])
dequeue_code = queue.dequeue_many(batch_size)
queue_runner = tf.train.queue_runner.QueueRunner(queue, [enqueue_op])
tf.add_to_collection(tf.GraphKeys.QUEUE_RUNNERS, queue_runner)
return dequeue_code
research/compression/entropy_coder/core/config_helper.py deleted 100644 → 0
View file @ 09bc9f54
# Copyright 2017 The TensorFlow Authors All Rights Reserved.
#
# 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.
# ==============================================================================
"""Helper functions used in both train and inference."""
import json
import os.path
import tensorflow as tf
def GetConfigString(config_file):
config_string = ''
if config_file is not None:
config_string = open(config_file).read()
return config_string
class InputConfig(object):
def __init__(self, config_string):
config = json.loads(config_string)
self.data = config["data"]
self.unique_code_size = config["unique_code_size"]
class TrainConfig(object):
def __init__(self, config_string):
config = json.loads(config_string)
self.batch_size = config["batch_size"]
self.learning_rate = config["learning_rate"]
self.decay_rate = config["decay_rate"]
self.samples_per_decay = config["samples_per_decay"]
def SaveConfig(directory, filename, config_string):
path = os.path.join(directory, filename)
with tf.gfile.Open(path, mode='w') as f:
f.write(config_string)
research/compression/entropy_coder/core/entropy_coder_single.py deleted 100644 → 0
View file @ 09bc9f54
# Copyright 2017 The TensorFlow Authors All Rights Reserved.
#
# 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.
# ==============================================================================
"""Compute the additional compression ratio after entropy coding."""
import io
import os
import numpy as np
import tensorflow as tf
import config_helper
# pylint: disable=unused-import
from entropy_coder.all_models import all_models
# pylint: enable=unused-import
from entropy_coder.model import model_factory
# Checkpoint used to restore the model parameters.
tf.app.flags.DEFINE_string('checkpoint', None,
"""Model checkpoint.""")
# Model selection and configuration.
tf.app.flags.DEFINE_string('model', None, """Underlying encoder model.""")
tf.app.flags.DEFINE_string('model_config', None,
"""Model config protobuf given as text file.""")
# File holding the binary codes.
tf.flags.DEFINE_string('input_codes', None, 'Location of binary code file.')
FLAGS = tf.flags.FLAGS
def main(_):
if (FLAGS.input_codes is None or FLAGS.model is None):
print ('\nUsage: python entropy_coder_single.py --model=progressive '
'--model_config=model_config.json'
'--iteration=15\n\n')
return
#if FLAGS.iteration < -1 or FLAGS.iteration > 15:
# print ('\n--iteration must be between 0 and 15 inclusive, or -1 to infer '
# 'from file.\n')
# return
#iteration = FLAGS.iteration
if not tf.gfile.Exists(FLAGS.input_codes):
print('\nInput codes not found.\n')
return
with tf.gfile.FastGFile(FLAGS.input_codes, 'rb') as code_file:
contents = code_file.read()
loaded_codes = np.load(io.BytesIO(contents))
assert ['codes', 'shape'] not in loaded_codes.files
loaded_shape = loaded_codes['shape']
loaded_array = loaded_codes['codes']
# Unpack and recover code shapes.
unpacked_codes = np.reshape(np.unpackbits(loaded_array)
[:np.prod(loaded_shape)],
loaded_shape)
numpy_int_codes = unpacked_codes.transpose([1, 2, 3, 0, 4])
numpy_int_codes = numpy_int_codes.reshape([numpy_int_codes.shape[0],
numpy_int_codes.shape[1],
numpy_int_codes.shape[2],
-1])
numpy_codes = numpy_int_codes.astype(np.float32) * 2.0 - 1.0
with tf.Graph().as_default() as graph:
# TF tensor to hold the binary codes to losslessly compress.
batch_size = 1
codes = tf.placeholder(tf.float32, shape=numpy_codes.shape)
# Create the entropy coder model.
global_step = None
optimizer = None
model = model_factory.GetModelRegistry().CreateModel(FLAGS.model)
model_config_string = config_helper.GetConfigString(FLAGS.model_config)
model.Initialize(global_step, optimizer, model_config_string)
model.BuildGraph(codes)
saver = tf.train.Saver(sharded=True, keep_checkpoint_every_n_hours=12.0)
with tf.Session(graph=graph) as sess:
# Initialize local variables.
sess.run(tf.local_variables_initializer())
# Restore model variables.
saver.restore(sess, FLAGS.checkpoint)
tf_tensors = {
'code_length': model.average_code_length
}
feed_dict = {codes: numpy_codes}
np_tensors = sess.run(tf_tensors, feed_dict=feed_dict)
print('Additional compression ratio: {}'.format(
np_tensors['code_length']))
if __name__ == '__main__':
tf.app.run()
research/compression/entropy_coder/core/entropy_coder_train.py deleted 100644 → 0
View file @ 09bc9f54
# Copyright 2017 The TensorFlow Authors All Rights Reserved.
#
# 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.
# ==============================================================================
"""Train an entropy coder model."""
import time
import tensorflow as tf
import code_loader
import config_helper
# pylint: disable=unused-import
from entropy_coder.all_models import all_models
# pylint: enable=unused-import
from entropy_coder.model import model_factory
FLAGS = tf.app.flags.FLAGS
# Hardware resources configuration.
tf.app.flags.DEFINE_string('master', '',
"""Name of the TensorFlow master to use.""")
tf.app.flags.DEFINE_string('train_dir', None,
"""Directory where to write event logs.""")
tf.app.flags.DEFINE_integer('task', None,
"""Task id of the replica running the training.""")
tf.app.flags.DEFINE_integer('ps_tasks', 0, """Number of tasks in the ps job.
If 0 no ps job is used.""")
# Model selection and configuration.
tf.app.flags.DEFINE_string('model', None, """Underlying encoder model.""")
tf.app.flags.DEFINE_string('model_config', None,
"""Model config protobuf given as text file.""")
# Training data and parameters configuration.
tf.app.flags.DEFINE_string('input_config', None,
"""Path to the training input config file.""")
tf.app.flags.DEFINE_string('train_config', None,
"""Path to the training experiment config file.""")
def train():
if FLAGS.train_dir is None:
raise ValueError('Parameter train_dir must be provided')
if FLAGS.task is None:
raise ValueError('Parameter task must be provided')
if FLAGS.model is None:
raise ValueError('Parameter model must be provided')
input_config_string = config_helper.GetConfigString(FLAGS.input_config)
input_config = config_helper.InputConfig(input_config_string)
# Training parameters.
train_config_string = config_helper.GetConfigString(FLAGS.train_config)
train_config = config_helper.TrainConfig(train_config_string)
batch_size = train_config.batch_size
initial_learning_rate = train_config.learning_rate
decay_rate = train_config.decay_rate
samples_per_decay = train_config.samples_per_decay
# Parameters for learning-rate decay.
# The formula is decay_rate ** floor(steps / decay_steps).
decay_steps = samples_per_decay / batch_size
decay_steps = max(decay_steps, 1)
first_code = code_loader.ReadFirstCode(input_config.data)
first_code_height = (
first_code.features.feature['code_shape'].int64_list.value[0])
first_code_width = (
first_code.features.feature['code_shape'].int64_list.value[1])
max_bit_depth = (
first_code.features.feature['code_shape'].int64_list.value[2])
print('Maximum code depth: {}'.format(max_bit_depth))
with tf.Graph().as_default():
ps_ops = ["Variable", "VariableV2", "AutoReloadVariable", "VarHandleOp"]
with tf.device(tf.train.replica_device_setter(FLAGS.ps_tasks,
ps_ops=ps_ops)):
codes = code_loader.LoadBinaryCode(
input_config=input_config,
batch_size=batch_size)
if input_config.unique_code_size:
print('Input code size: {} x {}'.format(first_code_height,
first_code_width))
codes.set_shape(
[batch_size, first_code_height, first_code_width, max_bit_depth])
else:
codes.set_shape([batch_size, None, None, max_bit_depth])
codes_effective_shape = tf.shape(codes)
global_step = tf.contrib.framework.create_global_step()
# Apply learning-rate decay.
learning_rate = tf.train.exponential_decay(
learning_rate=initial_learning_rate,
global_step=global_step,
decay_steps=decay_steps,
decay_rate=decay_rate,
staircase=True)
tf.summary.scalar('Learning Rate', learning_rate)
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate,
epsilon=1.0)
# Create the entropy coder model.
model = model_factory.GetModelRegistry().CreateModel(FLAGS.model)
model_config_string = config_helper.GetConfigString(FLAGS.model_config)
model.Initialize(global_step, optimizer, model_config_string)
model.BuildGraph(codes)
summary_op = tf.summary.merge_all()
# Verify that the model can actually be trained.
if model.train_op is None:
raise ValueError('Input model {} is not trainable'.format(FLAGS.model))
# We disable the summary thread run by Supervisor class by passing
# summary_op=None. We still pass save_summaries_secs because it is used by
# the global step counter thread.
is_chief = (FLAGS.task == 0)
sv = tf.train.Supervisor(logdir=FLAGS.train_dir,
is_chief=is_chief,
global_step=global_step,
# saver=model.saver,
summary_op=None,
save_summaries_secs=120,
save_model_secs=600,
recovery_wait_secs=30)
sess = sv.PrepareSession(FLAGS.master)
sv.StartQueueRunners(sess)
step = sess.run(global_step)
print('Trainer initial step: {}.'.format(step))
# Once everything has been setup properly, save the configs.
if is_chief:
config_helper.SaveConfig(FLAGS.train_dir, 'input_config.json',
input_config_string)
config_helper.SaveConfig(FLAGS.train_dir, 'model_config.json',
model_config_string)
config_helper.SaveConfig(FLAGS.train_dir, 'train_config.json',
train_config_string)
# Train the model.
next_summary_time = time.time()
while not sv.ShouldStop():
feed_dict = None
# Once in a while, update the summaries on the chief worker.
if is_chief and next_summary_time < time.time():
summary_str = sess.run(summary_op, feed_dict=feed_dict)
sv.SummaryComputed(sess, summary_str)
next_summary_time = time.time() + sv.save_summaries_secs
else:
tf_tensors = {
'train': model.train_op,
'code_length': model.average_code_length
}
np_tensors = sess.run(tf_tensors, feed_dict=feed_dict)
print(np_tensors['code_length'])
sv.Stop()
def main(argv=None): # pylint: disable=unused-argument
train()
if __name__ == '__main__':
tf.app.run()
research/compression/entropy_coder/dataset/gen_synthetic_dataset.py deleted 100644 → 0
View file @ 09bc9f54
# Copyright 2017 The TensorFlow Authors All Rights Reserved.
#
# 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.
# ==============================================================================
"""Generate a synthetic dataset."""
import os
import numpy as np
from six.moves import xrange
import tensorflow as tf
import synthetic_model
FLAGS = tf.app.flags.FLAGS
tf.app.flags.DEFINE_string(
'dataset_dir', None,
"""Directory where to write the dataset and the configs.""")
tf.app.flags.DEFINE_integer(
'count', 1000,
"""Number of samples to generate.""")
def int64_feature(values):
"""Returns a TF-Feature of int64s.
Args:
values: A scalar or list of values.
Returns:
A TF-Feature.
"""
if not isinstance(values, (tuple, list)):
values = [values]
return tf.train.Feature(int64_list=tf.train.Int64List(value=values))
def float_feature(values):
"""Returns a TF-Feature of floats.
Args:
values: A scalar of list of values.
Returns:
A TF-Feature.
"""
if not isinstance(values, (tuple, list)):
values = [values]
return tf.train.Feature(float_list=tf.train.FloatList(value=values))
def AddToTFRecord(code, tfrecord_writer):
example = tf.train.Example(features=tf.train.Features(feature={
'code_shape': int64_feature(code.shape),
'code': float_feature(code.flatten().tolist()),
}))
tfrecord_writer.write(example.SerializeToString())
def GenerateDataset(filename, count, code_shape):
with tf.python_io.TFRecordWriter(filename) as tfrecord_writer:
for _ in xrange(count):
code = synthetic_model.GenerateSingleCode(code_shape)
# Convert {0,1} codes to {-1,+1} codes.
code = 2.0 * code - 1.0
AddToTFRecord(code, tfrecord_writer)
def main(argv=None): # pylint: disable=unused-argument
GenerateDataset(os.path.join(FLAGS.dataset_dir + '/synthetic_dataset'),
FLAGS.count,
[35, 48, 8])
if __name__ == '__main__':
tf.app.run()
research/compression/entropy_coder/dataset/gen_synthetic_single.py deleted 100644 → 0
View file @ 09bc9f54
# Copyright 2016 The TensorFlow Authors All Rights Reserved.
#
# 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.
# ==============================================================================
"""Generate a single synthetic sample."""
import io
import os
import numpy as np
import tensorflow as tf
import synthetic_model
FLAGS = tf.app.flags.FLAGS
tf.app.flags.DEFINE_string(
'sample_filename', None,
"""Output file to store the generated binary code.""")
def GenerateSample(filename, code_shape, layer_depth):
# {0, +1} binary codes.
# No conversion since the output file is expected to store
# codes using {0, +1} codes (and not {-1, +1}).
code = synthetic_model.GenerateSingleCode(code_shape)
code = np.round(code)
# Reformat the code so as to be compatible with what is generated
# by the image encoder.
# The image encoder generates a tensor of size:
# iteration_count x batch_size x height x width x iteration_depth.
# Here: batch_size = 1
if code_shape[-1] % layer_depth != 0:
raise ValueError('Number of layers is not an integer')
height = code_shape[0]
width = code_shape[1]
code = code.reshape([1, height, width, -1, layer_depth])
code = np.transpose(code, [3, 0, 1, 2, 4])
int_codes = code.astype(np.int8)
exported_codes = np.packbits(int_codes.reshape(-1))
output = io.BytesIO()
np.savez_compressed(output, shape=int_codes.shape, codes=exported_codes)
with tf.gfile.FastGFile(filename, 'wb') as code_file:
code_file.write(output.getvalue())
def main(argv=None): # pylint: disable=unused-argument
# Note: the height and the width is different from the training dataset.
# The main purpose is to show that the entropy coder model is fully
# convolutional and can be used on any image size.
layer_depth = 2
GenerateSample(FLAGS.sample_filename, [31, 36, 8], layer_depth)
if __name__ == '__main__':
tf.app.run()
research/compression/entropy_coder/dataset/synthetic_model.py deleted 100644 → 0
View file @ 09bc9f54
# Copyright 2016 The TensorFlow Authors All Rights Reserved.
#
# 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.
# ==============================================================================
"""Binary code sample generator."""
import numpy as np
from six.moves import xrange
_CRC_LINE = [
[0, 1, 0],
[1, 1, 0],
[1, 0, 0]
]
_CRC_DEPTH = [1, 1, 0, 1]
def ComputeLineCrc(code, width, y, x, d):
crc = 0
for dy in xrange(len(_CRC_LINE)):
i = y - 1 - dy
if i < 0:
continue
for dx in xrange(len(_CRC_LINE[dy])):
j = x - 2 + dx
if j < 0 or j >= width:
continue
crc += 1 if (code[i, j, d] != _CRC_LINE[dy][dx]) else 0
return crc
def ComputeDepthCrc(code, y, x, d):
crc = 0
for delta in xrange(len(_CRC_DEPTH)):
k = d - 1 - delta
if k < 0:
continue
crc += 1 if (code[y, x, k] != _CRC_DEPTH[delta]) else 0
return crc
def GenerateSingleCode(code_shape):
code = np.zeros(code_shape, dtype=np.int)
keep_value_proba = 0.8
height = code_shape[0]
width = code_shape[1]
depth = code_shape[2]
for d in xrange(depth):
for y in xrange(height):
for x in xrange(width):
v1 = ComputeLineCrc(code, width, y, x, d)
v2 = ComputeDepthCrc(code, y, x, d)
v = 1 if (v1 + v2 >= 6) else 0
if np.random.rand() < keep_value_proba:
code[y, x, d] = v
else:
code[y, x, d] = 1 - v
return code
research/compression/entropy_coder/lib/block_base.py deleted 100644 → 0
View file @ 09bc9f54
# Copyright 2017 The TensorFlow Authors All Rights Reserved.
#
# 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.
# ==============================================================================
"""Base class for Tensorflow building blocks."""
import collections
import contextlib
import itertools
import tensorflow as tf
_block_stacks = collections.defaultdict(lambda: [])
class BlockBase(object):
"""Base class for transform wrappers of Tensorflow.
To implement a Tensorflow transform block, inherit this class.
1. To create a variable, use NewVar() method. Do not overload this method!
For example, use as follows.
a_variable = self.NewVar(initial_value)
2. All Tensorflow-related code must be done inside 'with self._BlockScope().'
Otherwise, name scoping and block hierarchy will not work. An exception
is _Apply() method, which is already called inside the context manager
by __call__() method.
3. Override and implement _Apply() method. This method is called by
__call__() method.
The users would use blocks like the following.
nn1 = NN(128, bias=Bias(0), act=tf.nn.relu)
y = nn1(x)
Some things to consider.
- Use lazy-initialization if possible. That is, initialize at first Apply()
rather than at __init__().
Note: if needed, the variables can be created on a specific parameter
server by creating blocks in a scope like:
with g.device(device):
linear = Linear(...)
"""
def __init__(self, name):
self._variables = []
self._subblocks = []
self._called = False
# Intentionally distinguishing empty string and None.
# If name is an empty string, then do not use name scope.
self.name = name if name is not None else self.__class__.__name__
self._graph = tf.get_default_graph()
if self.name:
# Capture the scope string at the init time.
with self._graph.name_scope(self.name) as scope:
self._scope_str = scope
else:
self._scope_str = ''
# Maintain hierarchy structure of blocks.
self._stack = _block_stacks[self._graph]
if self.__class__ is BlockBase:
# This code is only executed to create the root, which starts in the
# initialized state.
assert not self._stack
self._parent = None
self._called = True # The root is initialized.
return
# Create a fake root if a root is not already present.
if not self._stack:
self._stack.append(BlockBase('NoOpRoot'))
self._parent = self._stack[-1]
self._parent._subblocks.append(self) # pylint: disable=protected-access
def __repr__(self):
return '"{}" ({})'.format(self._scope_str, self.__class__.__name__)
@contextlib.contextmanager
def _OptionalNameScope(self, scope_str):
if scope_str:
with self._graph.name_scope(scope_str):
yield
else:
yield
@contextlib.contextmanager
def _BlockScope(self):
"""Context manager that handles graph, namescope, and nested blocks."""
self._stack.append(self)
try:
with self._graph.as_default():
with self._OptionalNameScope(self._scope_str):
yield self
finally: # Pop from the stack no matter exception is raised or not.
# The following line is executed when leaving 'with self._BlockScope()'
self._stack.pop()
def __call__(self, *args, **kwargs):
assert self._stack is _block_stacks[self._graph]
with self._BlockScope():
ret = self._Apply(*args, **kwargs)
self._called = True
return ret
def _Apply(self, *args, **kwargs):
"""Implementation of __call__()."""
raise NotImplementedError()
# Redirect all variable creation to this single function, so that we can
# switch to better variable creation scheme.
def NewVar(self, value, **kwargs):
"""Creates a new variable.
This function creates a variable, then returns a local copy created by
Identity operation. To get the Variable class object, use LookupRef()
method.
Note that each time Variable class object is used as an input to an
operation, Tensorflow will create a new Send/Recv pair. This hurts
performance.
If not for assign operations, use the local copy returned by this method.
Args:
value: Initialization value of the variable. The shape and the data type
of the variable is determined by this initial value.
**kwargs: Extra named arguments passed to Variable.__init__().
Returns:
A local copy of the new variable.
"""
v = tf.Variable(value, **kwargs)
self._variables.append(v)
return v
@property
def initialized(self):
"""Returns bool if the block is initialized.
By default, BlockBase assumes that a block is initialized when __call__()
is executed for the first time. If this is an incorrect assumption for some
subclasses, override this property in those subclasses.
Returns:
True if initialized, False otherwise.
"""
return self._called
def AssertInitialized(self):
"""Asserts initialized property."""
if not self.initialized:
raise RuntimeError('{} has not been initialized.'.format(self))
def VariableList(self):
"""Returns the list of all tensorflow variables used inside this block."""
variables = list(itertools.chain(
itertools.chain.from_iterable(
t.VariableList() for t in self._subblocks),
self._VariableList()))
return variables
def _VariableList(self):
"""Returns the list of all tensorflow variables owned by this block."""
self.AssertInitialized()
return self._variables
def CreateWeightLoss(self):
"""Returns L2 loss list of (almost) all variables used inside this block.
When this method needs to be overridden, there are two choices.
1. Override CreateWeightLoss() to change the weight loss of all variables
that belong to this block, both directly and indirectly.
2. Override _CreateWeightLoss() to change the weight loss of all
variables that directly belong to this block but not to the sub-blocks.
Returns:
A Tensor object or None.
"""
losses = list(itertools.chain(
itertools.chain.from_iterable(
t.CreateWeightLoss() for t in self._subblocks),
self._CreateWeightLoss()))
return losses
def _CreateWeightLoss(self):
"""Returns weight loss list of variables that belong to this block."""
self.AssertInitialized()
with self._BlockScope():
return [tf.nn.l2_loss(v) for v in self._variables]
def CreateUpdateOps(self):
"""Creates update operations for this block and its sub-blocks."""
ops = list(itertools.chain(
itertools.chain.from_iterable(
t.CreateUpdateOps() for t in self._subblocks),
self._CreateUpdateOps()))
return ops
def _CreateUpdateOps(self):
"""Creates update operations for this block."""
self.AssertInitialized()
return []
def MarkAsNonTrainable(self):
"""Mark all the variables of this block as non-trainable.
All the variables owned directly or indirectly (through subblocks) are
marked as non trainable.
This function along with CheckpointInitOp can be used to load a pretrained
model that consists in only one part of the whole graph.
"""
assert self._called
all_variables = self.VariableList()
collection = tf.get_collection_ref(tf.GraphKeys.TRAINABLE_VARIABLES)
for v in all_variables:
if v in collection:
collection.remove(v)
def CreateWeightLoss():
"""Returns all weight losses from the blocks in the graph."""
stack = _block_stacks[tf.get_default_graph()]
if not stack:
return []
return stack[0].CreateWeightLoss()
def CreateBlockUpdates():
"""Combines all updates from the blocks in the graph."""
stack = _block_stacks[tf.get_default_graph()]
if not stack:
return []
return stack[0].CreateUpdateOps()
research/compression/entropy_coder/lib/block_util.py deleted 100644 → 0
View file @ 09bc9f54
# Copyright 2017 The TensorFlow Authors All Rights Reserved.
#
# 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.
# ==============================================================================
"""Utility functions for blocks."""
from __future__ import division
from __future__ import unicode_literals
import math
import numpy as np
import six
import tensorflow as tf
class RsqrtInitializer(object):
"""Gaussian initializer with standard deviation 1/sqrt(n).
Note that tf.truncated_normal is used internally. Therefore any random sample
outside two-sigma will be discarded and re-sampled.
"""
def __init__(self, dims=(0,), **kwargs):
"""Creates an initializer.
Args:
dims: Dimension(s) index to compute standard deviation:
1.0 / sqrt(product(shape[dims]))
**kwargs: Extra keyword arguments to pass to tf.truncated_normal.
"""
if isinstance(dims, six.integer_types):
self._dims = [dims]
else:
self._dims = dims
self._kwargs = kwargs
def __call__(self, shape, dtype):
stddev = 1.0 / np.sqrt(np.prod([shape[x] for x in self._dims]))
return tf.truncated_normal(
shape=shape, dtype=dtype, stddev=stddev, **self._kwargs)
class RectifierInitializer(object):
"""Gaussian initializer with standard deviation sqrt(2/fan_in).
Note that tf.random_normal is used internally to ensure the expected weight
distribution. This is intended to be used with ReLU activations, specially
in ResNets.
For details please refer to:
Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet
Classification
"""
def __init__(self, dims=(0,), scale=2.0, **kwargs):
"""Creates an initializer.
Args:
dims: Dimension(s) index to compute standard deviation:
sqrt(scale / product(shape[dims]))
scale: A constant scaling for the initialization used as
sqrt(scale / product(shape[dims])).
**kwargs: Extra keyword arguments to pass to tf.truncated_normal.
"""
if isinstance(dims, six.integer_types):
self._dims = [dims]
else:
self._dims = dims
self._kwargs = kwargs
self._scale = scale
def __call__(self, shape, dtype):
stddev = np.sqrt(self._scale / np.prod([shape[x] for x in self._dims]))
return tf.random_normal(
shape=shape, dtype=dtype, stddev=stddev, **self._kwargs)
class GaussianInitializer(object):
"""Gaussian initializer with a given standard deviation.
Note that tf.truncated_normal is used internally. Therefore any random sample
outside two-sigma will be discarded and re-sampled.
"""
def __init__(self, stddev=1.0):
self._stddev = stddev
def __call__(self, shape, dtype):
return tf.truncated_normal(shape=shape, dtype=dtype, stddev=self._stddev)
research/compression/entropy_coder/lib/blocks.py deleted 100644 → 0
View file @ 09bc9f54
# Copyright 2017 The TensorFlow Authors All Rights Reserved.
#
# 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 block_base import *
from block_util import *
from blocks_binarizer import *
from blocks_entropy_coding import *
from blocks_lstm import *
from blocks_masked_conv2d import *
from blocks_masked_conv2d_lstm import *
from blocks_operator import *
from blocks_std import *
research/compression/entropy_coder/lib/blocks_binarizer.py deleted 100644 → 0
View file @ 09bc9f54
# Copyright 2017 The TensorFlow Authors All Rights Reserved.
#
# 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.
# ==============================================================================
"""Activation and weight binarizer implementations."""
import math
import numpy as np
import tensorflow as tf
def ConvertSignCodeToZeroOneCode(x):
"""Conversion from codes {-1, +1} to codes {0, 1}."""
return 0.5 * (x + 1.0)
def ConvertZeroOneCodeToSignCode(x):
"""Convert from codes {0, 1} to codes {-1, +1}."""
return 2.0 * x - 1.0
def CheckZeroOneCode(x):
return tf.reduce_all(tf.equal(x * (x - 1.0), 0))
research/compression/entropy_coder/lib/blocks_entropy_coding.py deleted 100644 → 0
View file @ 09bc9f54
# Copyright 2017 The TensorFlow Authors All Rights Reserved.
#
# 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.
# ==============================================================================
"""Set of blocks related to entropy coding."""
import math
import tensorflow as tf
import block_base
# pylint does not recognize block_base.BlockBase.__call__().
# pylint: disable=not-callable
class CodeLength(block_base.BlockBase):
"""Theoretical bound for a code length given a probability distribution.
"""
def __init__(self, name=None):
super(CodeLength, self).__init__(name)
def _Apply(self, c, p):
"""Theoretical bound of the coded length given a probability distribution.
Args:
c: The binary codes. Belong to {0, 1}.
p: The probability of: P(code==+1)
Returns:
The average code length.
Note: the average code length can be greater than 1 bit (e.g. when
encoding the least likely symbol).
"""
entropy = ((1.0 - c) * tf.log(1.0 - p) + c * tf.log(p)) / (-math.log(2))
entropy = tf.reduce_mean(entropy)
return entropy
research/compression/entropy_coder/lib/blocks_entropy_coding_test.py deleted 100644 → 0
View file @ 09bc9f54
# Copyright 2017 The TensorFlow Authors All Rights Reserved.
#
# 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.
# ==============================================================================
"""Tests for basic tensorflow blocks_entropy_coding."""
from __future__ import division
from __future__ import unicode_literals
import math
import numpy as np
import tensorflow as tf
import blocks_entropy_coding
class BlocksEntropyCodingTest(tf.test.TestCase):
def testCodeLength(self):
shape = [2, 4]
proba_feed = [[0.65, 0.25, 0.70, 0.10],
[0.28, 0.20, 0.44, 0.54]]
symbol_feed = [[1.0, 0.0, 1.0, 0.0],
[0.0, 0.0, 0.0, 1.0]]
mean_code_length = - (
(math.log(0.65) + math.log(0.75) + math.log(0.70) + math.log(0.90) +
math.log(0.72) + math.log(0.80) + math.log(0.56) + math.log(0.54)) /
math.log(2.0)) / (shape[0] * shape[1])
symbol = tf.placeholder(dtype=tf.float32, shape=shape)
proba = tf.placeholder(dtype=tf.float32, shape=shape)
code_length_calculator = blocks_entropy_coding.CodeLength()
code_length = code_length_calculator(symbol, proba)
with self.test_session():
tf.global_variables_initializer().run()
code_length_eval = code_length.eval(
feed_dict={symbol: symbol_feed, proba: proba_feed})
self.assertAllClose(mean_code_length, code_length_eval)
if __name__ == '__main__':
tf.test.main()
research/compression/entropy_coder/lib/blocks_lstm.py deleted 100644 → 0
View file @ 09bc9f54
# Copyright 2017 The TensorFlow Authors All Rights Reserved.
#
# 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.
# ==============================================================================
"""Blocks of LSTM and its variants."""
import numpy as np
import tensorflow as tf
import block_base
import block_util
import blocks_std
# pylint does not recognize block_base.BlockBase.__call__().
# pylint: disable=not-callable
def LSTMBiasInit(shape, dtype):
"""Returns ones for forget-gate, and zeros for the others."""
shape = np.array(shape)
# Check internal consistencies.
assert shape.shape == (1,), shape
assert shape[0] % 4 == 0, shape
n = shape[0] // 4
ones = tf.fill([n], tf.constant(1, dtype=dtype))
zeros = tf.fill([3 * n], tf.constant(0, dtype=dtype))
return tf.concat([ones, zeros], 0)
class LSTMBase(block_base.BlockBase):
"""Base class for LSTM implementations.
These LSTM implementations use the pattern found in [1]. No peephole
connection, i.e., cell content is not used in recurrence computation.
Hidden units are also output units.
[1] Zaremba, Sutskever, Vinyals. Recurrent Neural Network Regularization,
2015. arxiv:1409.2329.
"""
def __init__(self, output_shape, name):
"""Initializes LSTMBase class object.
Args:
output_shape: List representing the LSTM output shape. This argument
does not include batch dimension. For example, if the LSTM output has
shape [batch, depth], then pass [depth].
name: Name of this block.
"""
super(LSTMBase, self).__init__(name)
with self._BlockScope():
self._output_shape = [None] + list(output_shape)
self._hidden = None
self._cell = None
@property
def hidden(self):
"""Returns the hidden units of this LSTM."""
return self._hidden
@hidden.setter
def hidden(self, value):
"""Assigns to the hidden units of this LSTM.
Args:
value: The new value for the hidden units. If None, the hidden units are
considered to be filled with zeros.
"""
if value is not None:
value.get_shape().assert_is_compatible_with(self._output_shape)
self._hidden = value
@property
def cell(self):
"""Returns the cell units of this LSTM."""
return self._cell
@cell.setter
def cell(self, value):
"""Assigns to the cell units of this LSTM.
Args:
value: The new value for the cell units. If None, the cell units are
considered to be filled with zeros.
"""
if value is not None:
value.get_shape().assert_is_compatible_with(self._output_shape)
self._cell = value
# Consider moving bias terms to the base, and require this method to be
# linear.
def _TransformInputs(self, _):
"""Transforms the input units to (4 * depth) units.
The forget-gate, input-gate, output-gate, and cell update is computed as
f, i, j, o = T(h) + R(x)
where h is hidden units, x is input units, and T, R are transforms of
h, x, respectively.
This method implements R. Note that T is strictly linear, so if LSTM is
going to use bias, this method must include the bias to the transformation.
Subclasses must implement this method. See _Apply() for more details.
"""
raise NotImplementedError()
def _TransformHidden(self, _):
"""Transforms the hidden units to (4 * depth) units.
The forget-gate, input-gate, output-gate, and cell update is computed as
f, i, j, o = T(h) + R(x)
where h is hidden units, x is input units, and T, R are transforms of
h, x, respectively.
This method implements T in the equation. The method must implement a
strictly linear transformation. For example, it may use MatMul or Conv2D,
but must not add bias. This is because when hidden units are zeros, then
the LSTM implementation will skip calling this method, instead of passing
zeros to this function.
Subclasses must implement this method. See _Apply() for more details.
"""
raise NotImplementedError()
def _Apply(self, *args):
xtransform = self._TransformInputs(*args)
depth_axis = len(self._output_shape) - 1
if self.hidden is not None:
htransform = self._TransformHidden(self.hidden)
f, i, j, o = tf.split(
value=htransform + xtransform, num_or_size_splits=4, axis=depth_axis)
else:
f, i, j, o = tf.split(
value=xtransform, num_or_size_splits=4, axis=depth_axis)
if self.cell is not None:
self.cell = tf.sigmoid(f) * self.cell + tf.sigmoid(i) * tf.tanh(j)
else:
self.cell = tf.sigmoid(i) * tf.tanh(j)
self.hidden = tf.sigmoid(o) * tf.tanh(self.cell)
return self.hidden
class LSTM(LSTMBase):
"""Efficient LSTM implementation used in [1].
[1] Zaremba, Sutskever, Vinyals. Recurrent Neural Network Regularization,
2015. arxiv:1409.2329.
"""
def __init__(self,
depth,
bias=LSTMBiasInit,
initializer=block_util.RsqrtInitializer(),
name=None):
super(LSTM, self).__init__([depth], name)
with self._BlockScope():
self._depth = depth
self._nn = blocks_std.NN(
4 * depth, bias=bias, act=None, initializer=initializer)
self._hidden_linear = blocks_std.Linear(
4 * depth, initializer=initializer)
def _TransformInputs(self, *args):
return self._nn(*args)
def _TransformHidden(self, h):
return self._hidden_linear(h)
class Conv2DLSTM(LSTMBase):
"""Convolutional LSTM implementation with optimizations inspired by [1].
Note that when using the batch normalization feature, the bias initializer
will not be used, since BN effectively cancels its effect out.
[1] Zaremba, Sutskever, Vinyals. Recurrent Neural Network Regularization,
2015. arxiv:1409.2329.
"""
def __init__(self,
depth,
filter_size,
hidden_filter_size,
strides,
padding,
bias=LSTMBiasInit,
initializer=block_util.RsqrtInitializer(dims=(0, 1, 2)),
use_moving_average=False,
name=None):
super(Conv2DLSTM, self).__init__([None, None, depth], name)
self._iter = 0
with self._BlockScope():
self._input_conv = blocks_std.Conv2D(
4 * depth,
filter_size,
strides,
padding,
bias=None,
act=None,
initializer=initializer,
name='input_conv2d')
self._hidden_conv = blocks_std.Conv2D(
4 * depth,
hidden_filter_size,
[1, 1],
'SAME',
bias=None,
act=None,
initializer=initializer,
name='hidden_conv2d')
if bias is not None:
self._bias = blocks_std.BiasAdd(bias, name='biases')
else:
self._bias = blocks_std.PassThrough()
def _TransformInputs(self, x):
return self._bias(self._input_conv(x))
def _TransformHidden(self, h):
return self._hidden_conv(h)
def _Apply(self, *args):
xtransform = self._TransformInputs(*args)
depth_axis = len(self._output_shape) - 1
if self.hidden is not None:
htransform = self._TransformHidden(self.hidden)
f, i, j, o = tf.split(
value=htransform + xtransform, num_or_size_splits=4, axis=depth_axis)
else:
f, i, j, o = tf.split(
value=xtransform, num_or_size_splits=4, axis=depth_axis)
if self.cell is not None:
self.cell = tf.sigmoid(f) * self.cell + tf.sigmoid(i) * tf.tanh(j)
else:
self.cell = tf.sigmoid(i) * tf.tanh(j)
self.hidden = tf.sigmoid(o) * tf.tanh(self.cell)
self._iter += 1
return self.hidden
research/compression/entropy_coder/lib/blocks_lstm_test.py deleted 100644 → 0
View file @ 09bc9f54
# Copyright 2017 The TensorFlow Authors All Rights Reserved.
#
# 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.
# ==============================================================================
"""Tests for LSTM tensorflow blocks."""
from __future__ import division
import numpy as np
import tensorflow as tf
import block_base
import blocks_std
import blocks_lstm
class BlocksLSTMTest(tf.test.TestCase):
def CheckUnary(self, y, op_type):
self.assertEqual(op_type, y.op.type)
self.assertEqual(1, len(y.op.inputs))
return y.op.inputs[0]
def CheckBinary(self, y, op_type):
self.assertEqual(op_type, y.op.type)
self.assertEqual(2, len(y.op.inputs))
return y.op.inputs
def testLSTM(self):
lstm = blocks_lstm.LSTM(10)
lstm.hidden = tf.zeros(shape=[10, 10], dtype=tf.float32)
lstm.cell = tf.zeros(shape=[10, 10], dtype=tf.float32)
x = tf.placeholder(dtype=tf.float32, shape=[10, 11])
y = lstm(x)
o, tanhc = self.CheckBinary(y, 'Mul')
self.assertEqual(self.CheckUnary(o, 'Sigmoid').name, 'LSTM/split:3')
self.assertIs(lstm.cell, self.CheckUnary(tanhc, 'Tanh'))
fc, ij = self.CheckBinary(lstm.cell, 'Add')
f, _ = self.CheckBinary(fc, 'Mul')
self.assertEqual(self.CheckUnary(f, 'Sigmoid').name, 'LSTM/split:0')
i, j = self.CheckBinary(ij, 'Mul')
self.assertEqual(self.CheckUnary(i, 'Sigmoid').name, 'LSTM/split:1')
j = self.CheckUnary(j, 'Tanh')
self.assertEqual(j.name, 'LSTM/split:2')
def testLSTMBiasInit(self):
lstm = blocks_lstm.LSTM(9)
x = tf.placeholder(dtype=tf.float32, shape=[15, 7])
lstm(x)
b = lstm._nn._bias
with self.test_session():
tf.global_variables_initializer().run()
bias_var = b._bias.eval()
comp = ([1.0] * 9) + ([0.0] * 27)
self.assertAllEqual(bias_var, comp)
def testConv2DLSTM(self):
lstm = blocks_lstm.Conv2DLSTM(depth=10,
filter_size=[1, 1],
hidden_filter_size=[1, 1],
strides=[1, 1],
padding='SAME')
lstm.hidden = tf.zeros(shape=[10, 11, 11, 10], dtype=tf.float32)
lstm.cell = tf.zeros(shape=[10, 11, 11, 10], dtype=tf.float32)
x = tf.placeholder(dtype=tf.float32, shape=[10, 11, 11, 1])
y = lstm(x)
o, tanhc = self.CheckBinary(y, 'Mul')
self.assertEqual(self.CheckUnary(o, 'Sigmoid').name, 'Conv2DLSTM/split:3')
self.assertIs(lstm.cell, self.CheckUnary(tanhc, 'Tanh'))
fc, ij = self.CheckBinary(lstm.cell, 'Add')
f, _ = self.CheckBinary(fc, 'Mul')
self.assertEqual(self.CheckUnary(f, 'Sigmoid').name, 'Conv2DLSTM/split:0')
i, j = self.CheckBinary(ij, 'Mul')
self.assertEqual(self.CheckUnary(i, 'Sigmoid').name, 'Conv2DLSTM/split:1')
j = self.CheckUnary(j, 'Tanh')
self.assertEqual(j.name, 'Conv2DLSTM/split:2')
def testConv2DLSTMBiasInit(self):
lstm = blocks_lstm.Conv2DLSTM(9, 1, 1, [1, 1], 'SAME')
x = tf.placeholder(dtype=tf.float32, shape=[1, 7, 7, 7])
lstm(x)
b = lstm._bias
with self.test_session():
tf.global_variables_initializer().run()
bias_var = b._bias.eval()
comp = ([1.0] * 9) + ([0.0] * 27)
self.assertAllEqual(bias_var, comp)
if __name__ == '__main__':
tf.test.main()
research/compression/entropy_coder/lib/blocks_masked_conv2d.py deleted 100644 → 0
View file @ 09bc9f54
# Copyright 2017 The TensorFlow Authors All Rights Reserved.
#
# 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.
# ==============================================================================
"""Define some typical masked 2D convolutions."""
import numpy as np
from six.moves import xrange
import tensorflow as tf
import block_util
import blocks_std
# pylint does not recognize block_base.BlockBase.__call__().
# pylint: disable=not-callable
class RasterScanConv2D(blocks_std.Conv2DBase):
"""Conv2D with no dependency on future pixels (in raster scan order).
For example, assuming a 5 x 5 kernel, the kernel is applied a spatial mask:
T T T T T
T T T T T
T T x F F
F F F F F
F F F F F
where 'T' are pixels which are available when computing the convolution
for pixel 'x'. All the pixels marked with 'F' are not available.
'x' itself is not available if strict_order is True, otherwise, it is
available.
"""
def __init__(self, depth, filter_size, strides, padding,
strict_order=True,
bias=None, act=None, initializer=None, name=None):
super(RasterScanConv2D, self).__init__(
depth, filter_size, strides, padding, bias, act, name=name)
if (filter_size[0] % 2) != 1 or (filter_size[1] % 2) != 1:
raise ValueError('Kernel size should be odd.')
with self._BlockScope():
if initializer is None:
initializer = block_util.RsqrtInitializer(dims=(0, 1, 2))
self._initializer = initializer
self._strict_order = strict_order
def _CreateKernel(self, shape, dtype):
init = self._initializer(shape, dtype)
kernel = self.NewVar(init)
mask = np.ones(shape[:2], dtype=dtype.as_numpy_dtype)
center = shape[:2] // 2
mask[center[0] + 1:, :] = 0
if not self._strict_order:
mask[center[0], center[1] + 1:] = 0
else:
mask[center[0], center[1]:] = 0
mask = mask.reshape(mask.shape + (1, 1))
return tf.convert_to_tensor(mask, dtype) * kernel
class DepthOrderConv2D(blocks_std.Conv2DBase):
"""Conv2D with no dependency on higher depth dimensions.
More precisely, the output depth #n has only dependencies on input depths #k
for k < n (if strict_order is True) or for k <= n (if strict_order is False).
"""
def __init__(self, depth, filter_size, strides, padding,
strict_order=True,
bias=None, act=None, initializer=None, name=None):
super(DepthOrderConv2D, self).__init__(
depth, filter_size, strides, padding, bias, act, name=name)
with self._BlockScope():
if initializer is None:
initializer = block_util.RsqrtInitializer(dims=(0, 1, 2))
self._initializer = initializer
self._strict_order = strict_order
def _CreateKernel(self, shape, dtype):
init = self._initializer(shape, dtype)
kernel = self.NewVar(init)
mask = np.ones(shape[2:], dtype=dtype.as_numpy_dtype)
depth_output = shape[3]
for d in xrange(depth_output):
if self._strict_order:
mask[d:, d] = 0
else:
mask[d + 1:, d] = 0
mask = mask.reshape((1, 1) + mask.shape)
return tf.convert_to_tensor(mask, dtype) * kernel
class GroupRasterScanConv2D(blocks_std.Conv2DBase):
"""Conv2D with no dependency on future pixels (in raster scan order).
This version only introduces dependencies on previous pixels in raster scan
order. It can also introduce some dependencies on previous depth positions
of the current pixel (current pixel = center pixel of the kernel) in the
following way:
the depth dimension of the input is split into Ki groups of size
|input_group_size|, the output dimension is split into Ko groups of size
|output_group_size| (usually Ki == Ko). Each output group ko of the current
pixel position can only depend on previous input groups ki
(i.e. ki < ko if strict_order is True or ki <= ko if strict_order is False).
Notes:
- Block RasterScanConv2D is a special case of GroupRasterScanConv2D
where Ki == Ko == 1 (i.e. input_group_size == input_depth and
output_group_size == output_depth).
- For 1x1 convolution, block DepthOrderConv2D is a special case of
GroupRasterScanConv2D where input_group_size == 1 and
output_group_size == 1.
"""
def __init__(self, depth, filter_size, strides, padding,
strict_order=True,
input_group_size=1,
output_group_size=1,
bias=None, act=None, initializer=None, name=None):
super(GroupRasterScanConv2D, self).__init__(
depth, filter_size, strides, padding, bias, act, name=name)
if (filter_size[0] % 2) != 1 or (filter_size[1] % 2) != 1:
raise ValueError('Kernel size should be odd.')
with self._BlockScope():
if initializer is None:
initializer = block_util.RsqrtInitializer(dims=(0, 1, 2))
self._initializer = initializer
self._input_group_size = input_group_size
self._output_group_size = output_group_size
self._strict_order = strict_order
if depth % self._output_group_size != 0:
raise ValueError(
'Invalid depth group size: {} for depth {}'.format(
self._output_group_size, depth))
self._output_group_count = depth // self._output_group_size
def _CreateKernel(self, shape, dtype):
init = self._initializer(shape, dtype)
kernel = self.NewVar(init)
depth_input = shape[2]
if depth_input % self._input_group_size != 0:
raise ValueError(
'Invalid depth group size: {} for depth {}'.format(
self._input_group_size, depth_input))
input_group_count = depth_input // self._input_group_size
output_group_count = self._output_group_count
# Set the mask to 0 for future pixels in raster scan order.
center = shape[:2] // 2
mask = np.ones([shape[0], shape[1],
input_group_count, self._input_group_size,
output_group_count, self._output_group_size],
dtype=dtype.as_numpy_dtype)
mask[center[0] + 1:, :, :, :, :, :] = 0
mask[center[0], center[1] + 1:, :, :, :, :] = 0
# Adjust the mask for the current position (the center position).
depth_output = shape[3]
for d in xrange(output_group_count):
mask[center[0], center[1], d + 1:, :, d:d + 1, :] = 0
if self._strict_order:
mask[center[0], center[1], d, :, d:d + 1, :] = 0
mask = mask.reshape([shape[0], shape[1], depth_input, depth_output])
return tf.convert_to_tensor(mask, dtype) * kernel
class InFillingConv2D(blocks_std.Conv2DBase):
"""Conv2D with kernel having no dependency on the current pixel.
For example, assuming a 5 x 5 kernel, the kernel is applied a spatial mask:
T T T T T
T T T T T
T T x T T
T T T T T
T T T T T
where 'T' marks a pixel which is available when computing the convolution
for pixel 'x'. 'x' itself is not available.
"""
def __init__(self, depth, filter_size, strides, padding,
bias=None, act=None, initializer=None, name=None):
super(InFillingConv2D, self).__init__(
depth, filter_size, strides, padding, bias, act, name=name)
if (filter_size[0] % 2) != 1 or (filter_size[1] % 2) != 1:
raise ValueError('Kernel size should be odd.')
if filter_size[0] == 1 and filter_size[1] == 1:
raise ValueError('Kernel size should be larger than 1x1.')
with self._BlockScope():
if initializer is None:
initializer = block_util.RsqrtInitializer(dims=(0, 1, 2))
self._initializer = initializer
def _CreateKernel(self, shape, dtype):
init = self._initializer(shape, dtype)
kernel = self.NewVar(init)
mask = np.ones(shape[:2], dtype=dtype.as_numpy_dtype)
center = shape[:2] // 2
mask[center[0], center[1]] = 0
mask = mask.reshape(mask.shape + (1, 1))
return tf.convert_to_tensor(mask, dtype) * kernel
research/compression/entropy_coder/lib/blocks_masked_conv2d_lstm.py deleted 100644 → 0
View file @ 09bc9f54
# Copyright 2017 The TensorFlow Authors All Rights Reserved.
#
# 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.
# ==============================================================================
"""Masked conv2d LSTM."""
import block_base
import block_util
import blocks_masked_conv2d
import blocks_lstm
import blocks_std
# pylint: disable=not-callable
class RasterScanConv2DLSTM(blocks_lstm.LSTMBase):
"""Convolutional LSTM implementation with optimizations inspired by [1].
Note that when using the batch normalization feature, the bias initializer
will not be used, since BN effectively cancels its effect out.
[1] Zaremba, Sutskever, Vinyals. Recurrent Neural Network Regularization,
2015. arxiv:1409.2329.
"""
def __init__(self,
depth,
filter_size,
hidden_filter_size,
strides,
padding,
bias=blocks_lstm.LSTMBiasInit,
initializer=block_util.RsqrtInitializer(dims=(0, 1, 2)),
name=None):
super(RasterScanConv2DLSTM, self).__init__([None, None, depth], name)
with self._BlockScope():
self._input_conv = blocks_masked_conv2d.RasterScanConv2D(
4 * depth,
filter_size,
strides,
padding,
strict_order=False,
bias=None,
act=None,
initializer=initializer,
name='input_conv2d')
self._hidden_conv = blocks_std.Conv2D(
4 * depth,
hidden_filter_size,
[1, 1],
'SAME',
bias=None,
act=None,
initializer=initializer,
name='hidden_conv2d')
if bias is not None:
self._bias = blocks_std.BiasAdd(bias, name='biases')
else:
self._bias = blocks_std.PassThrough()
def _TransformInputs(self, x):
return self._bias(self._input_conv(x))
def _TransformHidden(self, h):
return self._hidden_conv(h)
research/compression/entropy_coder/lib/blocks_masked_conv2d_test.py deleted 100644 → 0
View file @ 09bc9f54
# Copyright 2017 The TensorFlow Authors All Rights Reserved.
#
# 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.
# ==============================================================================
"""Tests of the 2D masked convolution blocks."""
from __future__ import division
from __future__ import unicode_literals
import numpy as np
from six.moves import xrange
import tensorflow as tf
import blocks_masked_conv2d
class MaskedConv2DTest(tf.test.TestCase):
def testRasterScanKernel(self):
kernel_size = 5
input_depth = 1
output_depth = 1
kernel_shape = [kernel_size, kernel_size, input_depth, output_depth]
# pylint: disable=bad-whitespace
kernel_feed = [[ 1.0, 2.0, 3.0, 4.0, 5.0],
[ 6.0, 7.0, 8.0, 9.0, 10.0],
[11.0, 12.0, 13.0, 14.0, 15.0],
[16.0, 17.0, 18.0, 19.0, 20.0],
[21.0, 22.0, 23.0, 24.0, 25.0]]
kernel_feed = np.reshape(kernel_feed, kernel_shape)
kernel_expected = [[ 1.0, 2.0, 3.0, 4.0, 5.0],
[ 6.0, 7.0, 8.0, 9.0, 10.0],
[11.0, 12.0, 0.0, 0.0, 0.0],
[ 0.0, 0.0, 0.0, 0.0, 0.0],
[ 0.0, 0.0, 0.0, 0.0, 0.0]]
kernel_expected = np.reshape(kernel_expected, kernel_shape)
# pylint: enable=bad-whitespace
init_kernel = lambda s, t: tf.constant(kernel_feed, dtype=t, shape=s)
masked_conv2d = blocks_masked_conv2d.RasterScanConv2D(
output_depth, [kernel_size] * 2, [1] * 2, 'SAME',
initializer=init_kernel)
x = tf.placeholder(dtype=tf.float32, shape=[10] * 3 + [input_depth])
_ = masked_conv2d(x)
with self.test_session():
tf.global_variables_initializer().run()
kernel_value = masked_conv2d._kernel.eval()
self.assertAllEqual(kernel_expected, kernel_value)
def testDepthOrderKernel(self):
kernel_size = 1
input_depth = 7
output_depth = input_depth
kernel_shape = [kernel_size, kernel_size, input_depth, output_depth]
kernel_feed = np.ones(kernel_shape)
x_shape = [5] * 3 + [input_depth]
x_feed = np.ones(x_shape)
y_expected = np.zeros(x_shape[0:3] + [output_depth])
y_expected[:, :, :] = np.arange(output_depth)
init_kernel = lambda s, t: tf.constant(kernel_feed, dtype=t, shape=s)
masked_conv2d = blocks_masked_conv2d.DepthOrderConv2D(
output_depth, [kernel_size] * 2, [1] * 2, 'SAME',
strict_order=True,
initializer=init_kernel)
x = tf.placeholder(dtype=tf.float32, shape=x_shape)
y = masked_conv2d(x)
with self.test_session():
tf.global_variables_initializer().run()
y_value = y.eval(feed_dict={x: x_feed})
self.assertAllEqual(y_expected, y_value)
def testGroupRasterScanKernel(self):
kernel_size = 3
input_depth = 4
input_group_size = 2
output_depth = 2
output_group_size = 1
kernel_shape = [kernel_size, kernel_size, input_depth, output_depth]
kernel_feed = np.ones(shape=kernel_shape)
height = 5
width = 5
x_shape = [1, height, width, input_depth]
x_feed = np.ones(shape=x_shape)
# pylint: disable=bad-whitespace
y_expected = [
[[ 0, 2], [ 4, 6], [ 4, 6], [ 4, 6], [ 4, 6]],
[[ 8, 10], [16, 18], [16, 18], [16, 18], [12, 14]],
[[ 8, 10], [16, 18], [16, 18], [16, 18], [12, 14]],
[[ 8, 10], [16, 18], [16, 18], [16, 18], [12, 14]],
[[ 8, 10], [16, 18], [16, 18], [16, 18], [12, 14]],
]
y_expected = np.reshape(y_expected, [1, height, width, output_depth])
# pylint: enable=bad-whitespace
init_kernel = lambda s, t: tf.constant(kernel_feed, dtype=t, shape=s)
masked_conv2d = blocks_masked_conv2d.GroupRasterScanConv2D(
output_depth, [kernel_size] * 2, [1] * 2, 'SAME',
strict_order=True,
input_group_size=input_group_size,
output_group_size=output_group_size,
initializer=init_kernel)
x = tf.placeholder(dtype=tf.float32, shape=x_shape)
y = masked_conv2d(x)
with self.test_session():
tf.global_variables_initializer().run()
y_value = y.eval(feed_dict={x: x_feed})
self.assertAllEqual(y_expected, y_value)
def testInFillingKernel(self):
kernel_size = 5
input_depth = 1
output_depth = 1
kernel_shape = [kernel_size, kernel_size, input_depth, output_depth]
# pylint: disable=bad-whitespace
kernel_feed = [[ 1.0, 2.0, 3.0, 4.0, 5.0],
[ 6.0, 7.0, 8.0, 9.0, 10.0],
[11.0, 12.0, 13.0, 14.0, 15.0],
[16.0, 17.0, 18.0, 19.0, 20.0],
[21.0, 22.0, 23.0, 24.0, 25.0]]
kernel_feed = np.reshape(kernel_feed, kernel_shape)
kernel_expected = [[ 1.0, 2.0, 3.0, 4.0, 5.0],
[ 6.0, 7.0, 8.0, 9.0, 10.0],
[11.0, 12.0, 0.0, 14.0, 15.0],
[16.0, 17.0, 18.0, 19.0, 20.0],
[21.0, 22.0, 23.0, 24.0, 25.0]]
kernel_expected = np.reshape(kernel_expected, kernel_shape)
# pylint: enable=bad-whitespace
init_kernel = lambda s, t: tf.constant(kernel_feed, dtype=t, shape=s)
masked_conv2d = blocks_masked_conv2d.InFillingConv2D(
output_depth, [kernel_size] * 2, [1] * 2, 'SAME',
initializer=init_kernel)
x = tf.placeholder(dtype=tf.float32, shape=[10] * 3 + [input_depth])
_ = masked_conv2d(x)
with self.test_session():
tf.global_variables_initializer().run()
kernel_value = masked_conv2d._kernel.eval()
self.assertAllEqual(kernel_expected, kernel_value)
def testConv2DMaskedNumerics(self):
kernel_size = 5
input_shape = [1, 10, 10, 1]
filter_shape = [kernel_size, kernel_size, 1, 1]
strides = [1, 1, 1, 1]
output_shape = [1, 10, 10, 1]
conv = blocks_masked_conv2d.RasterScanConv2D(
depth=filter_shape[-1],
filter_size=filter_shape[0:2],
strides=strides[1:3],
padding='SAME',
initializer=tf.constant_initializer(value=1.0))
x = tf.placeholder(dtype=tf.float32, shape=input_shape)
y = conv(x)
x_feed = - np.ones(input_shape, dtype=float)
y_expected = np.ones(output_shape, dtype=float)
for i in xrange(input_shape[1]):
for j in xrange(input_shape[2]):
x_feed[0, i, j, 0] = 10 * (j + 1) + i
v = 0
ki_start = max(i - kernel_size // 2, 0)
kj_start = max(j - kernel_size // 2, 0)
kj_end = min(j + kernel_size // 2, input_shape[2] - 1)
for ki in range(ki_start, i + 1):
for kj in range(kj_start, kj_end + 1):
if ki > i:
continue
if ki == i and kj >= j:
continue
v += 10 * (kj + 1) + ki
y_expected[0, i, j, 0] = v
with self.test_session():
tf.global_variables_initializer().run()
y_value = y.eval(feed_dict={x: x_feed})
self.assertAllEqual(y_expected, y_value)
if __name__ == '__main__':
tf.test.main()
research/compression/entropy_coder/lib/blocks_operator.py deleted 100644 → 0
View file @ 09bc9f54
# Copyright 2017 The TensorFlow Authors All Rights Reserved.
#
# 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.
# ==============================================================================
"""Common blocks which work as operators on other blocks."""
import tensorflow as tf
import block_base
# pylint: disable=not-callable
class CompositionOperator(block_base.BlockBase):
"""Composition of several blocks."""
def __init__(self, block_list, name=None):
"""Initialization of the composition operator.
Args:
block_list: List of blocks.BlockBase that are chained to create
a new blocks.BlockBase.
name: Name of this block.
"""
super(CompositionOperator, self).__init__(name)
self._blocks = block_list
def _Apply(self, x):
"""Apply successively all the blocks on the given input tensor."""
h = x
for layer in self._blocks:
h = layer(h)
return h
class LineOperator(block_base.BlockBase):
"""Repeat the same block over all the lines of an input tensor."""
def __init__(self, block, name=None):
super(LineOperator, self).__init__(name)
self._block = block
def _Apply(self, x):
height = x.get_shape()[1].value
if height is None:
raise ValueError('Unknown tensor height')
all_line_x = tf.split(value=x, num_or_size_splits=height, axis=1)
y = []
for line_x in all_line_x:
y.append(self._block(line_x))
y = tf.concat(values=y, axis=1)
return y
class TowerOperator(block_base.BlockBase):
"""Parallel execution with concatenation of several blocks."""
def __init__(self, block_list, dim=3, name=None):
"""Initialization of the parallel exec + concat (Tower).
Args:
block_list: List of blocks.BlockBase that are chained to create
a new blocks.BlockBase.
dim: the dimension on which to concat.
name: Name of this block.
"""
super(TowerOperator, self).__init__(name)
self._blocks = block_list
self._concat_dim = dim
def _Apply(self, x):
"""Apply successively all the blocks on the given input tensor."""
outputs = [layer(x) for layer in self._blocks]
return tf.concat(outputs, self._concat_dim)
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment