Merge pull request #1499 from damienv-gh/master

Image compression: initial version of the entropy coder.

compression/entropy_coder/lib/blocks_binarizer.py

+# 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))

compression/entropy_coder/lib/blocks_entropy_coding.py

+# 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

compression/entropy_coder/lib/blocks_entropy_coding_test.py

+# 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()

compression/entropy_coder/lib/blocks_lstm.py

+# 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

compression/entropy_coder/lib/blocks_lstm_test.py

+# 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()

compression/entropy_coder/lib/blocks_masked_conv2d.py

+# 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
+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

compression/entropy_coder/lib/blocks_masked_conv2d_lstm.py

+# 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)

compression/entropy_coder/lib/blocks_masked_conv2d_test.py

+# 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
+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()

compression/entropy_coder/lib/blocks_operator.py

+# 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)

compression/entropy_coder/lib/blocks_operator_test.py

+# 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 block operators."""
+
+import numpy as np
+import tensorflow as tf
+
+import block_base
+import blocks_operator
+
+
+class AddOneBlock(block_base.BlockBase):
+
+  def __init__(self, name=None):
+    super(AddOneBlock, self).__init__(name)
+
+  def _Apply(self, x):
+    return x + 1.0
+
+
+class SquareBlock(block_base.BlockBase):
+
+  def __init__(self, name=None):
+    super(SquareBlock, self).__init__(name)
+
+  def _Apply(self, x):
+    return x * x
+
+
+class BlocksOperatorTest(tf.test.TestCase):
+
+  def testComposition(self):
+    x_value = np.array([[1.0, 2.0, 3.0],
+                        [-1.0, -2.0, -3.0]])
+    y_expected_value = np.array([[4.0, 9.0, 16.0],
+                                 [0.0, 1.0, 4.0]])
+
+    x = tf.placeholder(dtype=tf.float32, shape=[2, 3])
+    complex_block = blocks_operator.CompositionOperator(
+        [AddOneBlock(),
+         SquareBlock()])
+    y = complex_block(x)
+
+    with self.test_session():
+      y_value = y.eval(feed_dict={x: x_value})
+
+    self.assertAllClose(y_expected_value, y_value)
+
+
+if __name__ == '__main__':
+  tf.test.main()

compression/entropy_coder/lib/blocks_std.py

+# 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.
+# ==============================================================================
+
+"""Basic blocks for building tensorflow models."""
+
+import numpy as np
+import tensorflow as tf
+
+import block_base
+import block_util
+
+# pylint does not recognize block_base.BlockBase.__call__().
+# pylint: disable=not-callable
+
+
+def HandleConvPaddingModes(x, padding, kernel_shape, strides):
+  """Returns an updated tensor and padding type for REFLECT and SYMMETRIC.
+
+  Args:
+    x: A 4D tensor with shape [batch_size, height, width, depth].
+    padding: Padding mode (SAME, VALID, REFLECT, or SYMMETRIC).
+    kernel_shape: Shape of convolution kernel that will be applied.
+    strides: Convolution stride that will be used.
+
+  Returns:
+    x and padding after adjustments for REFLECT and SYMMETRIC.
+  """
+  # For 1x1 convolution, all padding modes are the same.
+  if np.all(kernel_shape[:2] == 1):
+    return x, 'VALID'
+
+  if padding == 'REFLECT' or padding == 'SYMMETRIC':
+    # We manually compute the number of paddings as if 'SAME'.
+    # From Tensorflow kernel, the formulas are as follows.
+    #   output_shape = ceil(input_shape / strides)
+    #   paddings = (output_shape - 1) * strides + filter_size - input_shape
+    # Let x, y, s be a shorthand notations for input_shape, output_shape, and
+    # strides, respectively. Let (x - 1) = sn + r where 0 <= r < s. Note that
+    #   y - 1 = ceil(x / s) - 1 = floor((x - 1) / s) = n
+    # provided that x > 0. Therefore
+    #   paddings = n * s + filter_size - (sn + r + 1)
+    #            = filter_size - r - 1.
+    input_shape = x.get_shape()  # shape at graph construction time
+    img_shape = tf.shape(x)[1:3]  # image shape (no batch) at run time
+    remainder = tf.mod(img_shape - 1, strides[1:3])
+    pad_sizes = kernel_shape[:2] - remainder - 1
+
+    pad_rows = pad_sizes[0]
+    pad_cols = pad_sizes[1]
+    pad = tf.stack([[0, 0], tf.stack([pad_rows // 2, (pad_rows + 1) // 2]),
+                    tf.stack([pad_cols // 2, (pad_cols + 1) // 2]), [0, 0]])
+
+    # Manually pad the input and switch the padding mode to 'VALID'.
+    x = tf.pad(x, pad, mode=padding)
+    x.set_shape([input_shape[0], x.get_shape()[1],
+                 x.get_shape()[2], input_shape[3]])
+    padding = 'VALID'
+
+  return x, padding
+
+
+class PassThrough(block_base.BlockBase):
+  """A dummy transform block that does nothing."""
+
+  def __init__(self):
+    # Pass an empty string to disable name scoping.
+    super(PassThrough, self).__init__(name='')
+
+  def _Apply(self, inp):
+    return inp
+
+  @property
+  def initialized(self):
+    """Always returns True."""
+    return True
+
+
+class Bias(object):
+  """An initialization helper class for BiasAdd block below."""
+
+  def __init__(self, value=0):
+    self.value = value
+
+
+class BiasAdd(block_base.BlockBase):
+  """A tf.nn.bias_add wrapper.
+
+  This wrapper may act as a PassThrough block depending on the initializer
+  provided, to make easier optional bias applications in NN blocks, etc.
+  See __init__() for the details.
+  """
+
+  def __init__(self, initializer=Bias(0), name=None):
+    """Initializes Bias block.
+
+    |initializer| parameter have two special cases.
+
+    1. If initializer is None, then this block works as a PassThrough.
+    2. If initializer is a Bias class object, then tf.constant_initializer is
+       used with the stored value.
+
+    Args:
+      initializer: An initializer for the bias variable.
+      name: Name of this block.
+    """
+    super(BiasAdd, self).__init__(name)
+
+    with self._BlockScope():
+      if isinstance(initializer, Bias):
+        self._initializer = tf.constant_initializer(value=initializer.value)
+      else:
+        self._initializer = initializer
+
+      self._bias = None
+
+  def _Apply(self, x):
+    if not self._bias:
+      init = self._initializer([int(x.get_shape()[-1])], x.dtype)
+      self._bias = self.NewVar(init)
+
+    return tf.nn.bias_add(x, self._bias)
+
+  def CreateWeightLoss(self):
+    return []
+
+
+class LinearBase(block_base.BlockBase):
+  """A matmul wrapper.
+
+  Returns input * W, where matrix W can be customized through derivation.
+  """
+
+  def __init__(self, depth, name=None):
+    super(LinearBase, self).__init__(name)
+
+    with self._BlockScope():
+      self._depth = depth
+      self._matrix = None
+
+  def _CreateKernel(self, shape, dtype):
+    raise NotImplementedError('This method must be sub-classed.')
+
+  def _Apply(self, x):
+    if not self._matrix:
+      shape = [int(x.get_shape()[-1]), self._depth]
+      self._matrix = self._CreateKernel(shape, x.dtype)
+
+    return tf.matmul(x, self._matrix)
+
+
+class Linear(LinearBase):
+  """A matmul wrapper.
+
+  Returns input * W, where matrix W is learned.
+  """
+
+  def __init__(self,
+               depth,
+               initializer=block_util.RsqrtInitializer(),
+               name=None):
+    super(Linear, self).__init__(depth, name)
+
+    with self._BlockScope():
+      self._initializer = initializer
+
+  def _CreateKernel(self, shape, dtype):
+    init = self._initializer(shape, dtype)
+    return self.NewVar(init)
+
+
+class NN(block_base.BlockBase):
+  """A neural network layer wrapper.
+
+  Returns act(input * W + b), where matrix W, bias b are learned, and act is an
+  optional activation function (i.e., nonlinearity).
+
+  This transform block can handle multiple inputs. If x_1, x_2, ..., x_m are
+  the inputs, then returns act(x_1 * W_1 + ... + x_m * W_m + b).
+
+  Attributes:
+    nunits: The dimension of the output.
+  """
+
+  def __init__(self,
+               depth,
+               bias=Bias(0),
+               act=None,  # e.g., tf.nn.relu
+               initializer=block_util.RsqrtInitializer(),
+               linear_block_factory=(lambda d, i: Linear(d, initializer=i)),
+               name=None):
+    """Initializes NN block.
+
+    Args:
+      depth: The depth of the output.
+      bias: An initializer for the bias, or a Bias class object. If None, there
+        will be no bias term for this NN block. See BiasAdd block.
+      act: Optional activation function. If None, no activation is applied.
+      initializer: The initialization method for the matrix weights.
+      linear_block_factory: A function used to create a linear block.
+      name: The name of this block.
+    """
+    super(NN, self).__init__(name)
+
+    with self._BlockScope():
+      self._linear_block_factory = linear_block_factory
+      self._depth = depth
+      self._initializer = initializer
+      self._matrices = None
+
+      self._bias = BiasAdd(bias) if bias else PassThrough()
+      self._act = act if act else PassThrough()
+
+  def _Apply(self, *args):
+    if not self._matrices:
+      self._matrices = [
+          self._linear_block_factory(self._depth, self._initializer)
+          for _ in args]
+
+    if len(self._matrices) != len(args):
+      raise ValueError('{} expected {} inputs, but observed {} inputs'.format(
+          self.name, len(self._matrices), len(args)))
+
+    if len(args) > 1:
+      y = tf.add_n([m(x) for m, x in zip(self._matrices, args)])
+    else:
+      y = self._matrices[0](args[0])
+
+    return self._act(self._bias(y))
+
+
+class Conv2DBase(block_base.BlockBase):
+  """A tf.nn.conv2d operator."""
+
+  def __init__(self, depth, filter_size, strides, padding,
+               bias=None, act=None, atrous_rate=None, conv=tf.nn.conv2d,
+               name=None):
+    """Initializes a Conv2DBase block.
+
+    Arguments:
+      depth: The output depth of the block (i.e. #filters); if negative, the
+        output depth will be set to be the same as the input depth.
+      filter_size: The size of the 2D filter. If it's specified as an integer,
+        it's going to create a square filter. Otherwise, this is a tuple
+        specifying the height x width of the filter.
+      strides: A tuple specifying the y and x stride.
+      padding: One of the valid padding modes allowed by tf.nn.conv2d, or
+        'REFLECT'/'SYMMETRIC' for mirror padding.
+      bias: An initializer for the bias, or a Bias class object. If None, there
+          will be no bias in this block. See BiasAdd block.
+      act: Optional activation function applied to the output.
+      atrous_rate: optional input rate for ATrous convolution. If not None, this
+          will be used and the strides will be ignored.
+      conv: The convolution function to use (e.g. tf.nn.conv2d).
+      name: The name for this conv2d op.
+    """
+    super(Conv2DBase, self).__init__(name)
+
+    with self._BlockScope():
+      self._act = act if act else PassThrough()
+      self._bias = BiasAdd(bias) if bias else PassThrough()
+
+      self._kernel_shape = np.zeros((4,), dtype=np.int32)
+      self._kernel_shape[:2] = filter_size
+      self._kernel_shape[3] = depth
+
+      self._strides = np.ones((4,), dtype=np.int32)
+      self._strides[1:3] = strides
+      self._strides = list(self._strides)
+
+      self._padding = padding
+
+      self._kernel = None
+      self._conv = conv
+
+      self._atrous_rate = atrous_rate
+
+  def _CreateKernel(self, shape, dtype):
+    raise NotImplementedError('This method must be sub-classed')
+
+  def _Apply(self, x):
+    """Apply the self._conv op.
+
+    Arguments:
+      x: input tensor. It needs to be a 4D tensor of the form
+          [batch, height, width, channels].
+    Returns:
+      The output of the convolution of x with the current convolutional
+      kernel.
+    Raises:
+      ValueError: if number of channels is not defined at graph construction.
+    """
+    input_shape = x.get_shape().with_rank(4)
+    input_shape[3:].assert_is_fully_defined()  # channels must be defined
+    if self._kernel is None:
+      assert self._kernel_shape[2] == 0, self._kernel_shape
+      self._kernel_shape[2] = input_shape[3].value
+      if self._kernel_shape[3] < 0:
+        # Make output depth be the same as input depth.
+        self._kernel_shape[3] = self._kernel_shape[2]
+      self._kernel = self._CreateKernel(self._kernel_shape, x.dtype)
+
+    x, padding = HandleConvPaddingModes(
+        x, self._padding, self._kernel_shape, self._strides)
+    if self._atrous_rate is None:
+      x = self._conv(x, self._kernel, strides=self._strides, padding=padding)
+    else:
+      x = self._conv(x, self._kernel, rate=self._atrous_rate, padding=padding)
+
+    if self._padding != 'VALID':
+      # Manually update shape. Known shape information can be lost by tf.pad().
+      height = (1 + (input_shape[1].value - 1) // self._strides[1]
+                if input_shape[1].value else None)
+      width = (1 + (input_shape[2].value - 1) // self._strides[2]
+               if input_shape[2].value else None)
+      shape = x.get_shape()
+      x.set_shape([shape[0], height, width, shape[3]])
+
+    return self._act(self._bias(x))
+
+
+class Conv2D(Conv2DBase):
+  """A tf.nn.conv2d operator."""
+
+  def __init__(self, depth, filter_size, strides, padding,
+               bias=None, act=None, initializer=None, name=None):
+    """Initializes a Conv2D block.
+
+    Arguments:
+      depth: The output depth of the block (i.e., #filters)
+      filter_size: The size of the 2D filter. If it's specified as an integer,
+        it's going to create a square filter. Otherwise, this is a tuple
+        specifying the height x width of the filter.
+      strides: A tuple specifying the y and x stride.
+      padding: One of the valid padding modes allowed by tf.nn.conv2d, or
+        'REFLECT'/'SYMMETRIC' for mirror padding.
+      bias: An initializer for the bias, or a Bias class object. If None, there
+          will be no bias in this block. See BiasAdd block.
+      act: Optional activation function applied to the output.
+      initializer: Optional initializer for weights.
+      name: The name for this conv2d op.
+    """
+    super(Conv2D, self).__init__(depth, filter_size, strides, padding, bias,
+                                 act, conv=tf.nn.conv2d, name=name)
+
+    with self._BlockScope():
+      if initializer is None:
+        initializer = block_util.RsqrtInitializer(dims=(0, 1, 2))
+      self._initializer = initializer
+
+  def _CreateKernel(self, shape, dtype):
+    return self.NewVar(self._initializer(shape, dtype))

compression/entropy_coder/lib/blocks_std_test.py

+# 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_std."""
+
+from __future__ import division
+from __future__ import unicode_literals
+
+import math
+import os
+
+import numpy as np
+import tensorflow as tf
+
+import blocks_std
+
+
+def _NumpyConv2D(x, f, strides, padding, rate=1):
+  assert strides[0] == 1 and strides[3] == 1, strides
+
+  if rate > 1:
+    f_shape = f.shape
+    expand_f = np.zeros([f_shape[0], ((f_shape[1] - 1) * rate + 1),
+                         f_shape[2], f_shape[3]])
+    expand_f[:, [y * rate for y in range(f_shape[1])], :, :] = f
+    f = np.zeros([((f_shape[0] - 1) * rate + 1), expand_f.shape[1],
+                  f_shape[2], f_shape[3]])
+    f[[y * rate for y in range(f_shape[0])], :, :, :] = expand_f
+
+  if padding != 'VALID':
+    assert x.shape[1] > 0 and x.shape[2] > 0, x.shape
+    # Compute the number of padded rows and cols.
+    # See Conv2D block comments for a math explanation.
+    remainder = ((x.shape[1] - 1) % strides[1], (x.shape[2] - 1) % strides[2])
+    pad_rows = f.shape[0] - remainder[0] - 1
+    pad_cols = f.shape[1] - remainder[1] - 1
+    pad = ((0, 0),
+           (pad_rows // 2, (pad_rows + 1) // 2),
+           (pad_cols // 2, (pad_cols + 1) // 2),
+           (0, 0))
+
+    # Pad the input using numpy.pad().
+    mode = None
+    if padding == 'SAME':
+      mode = str('constant')
+    if padding == 'REFLECT':
+      mode = str('reflect')
+    if padding == 'SYMMETRIC':
+      mode = str('symmetric')
+    x = np.pad(x, pad, mode=mode)
+
+  # Since x is now properly padded, proceed as if padding mode is VALID.
+  x_window = np.empty(
+      (x.shape[0],
+       int(math.ceil((x.shape[1] - f.shape[0] + 1) / strides[1])),
+       int(math.ceil((x.shape[2] - f.shape[1] + 1) / strides[2])),
+       np.prod(f.shape[:3])))
+
+  # The output at pixel location (i, j) is the result of linear transformation
+  # applied to the window whose top-left corner is at
+  # (i * row_stride, j * col_stride).
+  for i in xrange(x_window.shape[1]):
+    k = i * strides[1]
+    for j in xrange(x_window.shape[2]):
+      l = j * strides[2]
+      x_window[:, i, j, :] = x[:,
+                               k:(k + f.shape[0]),
+                               l:(l + f.shape[1]),
+                               :].reshape((x_window.shape[0], -1))
+
+  y = np.tensordot(x_window, f.reshape((-1, f.shape[3])), axes=1)
+  return y
+
+
+class BlocksStdTest(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 testPassThrough(self):
+    p = blocks_std.PassThrough()
+    x = tf.placeholder(dtype=tf.float32, shape=[1])
+    self.assertIs(p(x), x)
+
+  def CheckBiasAdd(self, y, b):
+    x, u = self.CheckBinary(y, 'BiasAdd')
+    self.assertIs(u, b._bias.value())
+    self.assertEqual(x.dtype, u.dtype.base_dtype)
+    return x
+
+  def testBiasAdd(self):
+    b = blocks_std.BiasAdd()
+    x = tf.placeholder(dtype=tf.float32, shape=[4, 8])
+    y = b(x)
+    self.assertEqual(b._bias.get_shape(), x.get_shape()[-1:])
+    self.assertIs(x, self.CheckBiasAdd(y, b))
+
+  def testBiasRankTest(self):
+    b = blocks_std.BiasAdd()
+    x = tf.placeholder(dtype=tf.float32, shape=[10])
+    with self.assertRaises(ValueError):
+      b(x)
+
+  def CheckLinear(self, y, m):
+    x, w = self.CheckBinary(y, 'MatMul')
+    self.assertIs(w, m._matrix.value())
+    self.assertEqual(x.dtype, w.dtype.base_dtype)
+    return x
+
+  def testLinear(self):
+    m = blocks_std.Linear(10)
+    x = tf.placeholder(dtype=tf.float32, shape=[8, 9])
+    y = m(x)
+    self.assertEqual(m._matrix.get_shape(), [9, 10])
+    self.assertIs(x, self.CheckLinear(y, m))
+
+  def testLinearShared(self):
+    # Create a linear map which is applied twice on different inputs
+    # (i.e. the weights of the map are shared).
+    linear_map = blocks_std.Linear(6)
+    x1 = tf.random_normal(shape=[1, 5])
+    x2 = tf.random_normal(shape=[1, 5])
+    xs = x1 + x2
+
+    # Apply the transform with the same weights.
+    y1 = linear_map(x1)
+    y2 = linear_map(x2)
+    ys = linear_map(xs)
+
+    with self.test_session() as sess:
+      # Initialize all the variables of the graph.
+      tf.global_variables_initializer().run()
+
+      y1_res, y2_res, ys_res = sess.run([y1, y2, ys])
+      self.assertAllClose(y1_res + y2_res, ys_res)
+
+  def CheckNN(self, y, nn, act=None):
+    if act:
+      pre_act = self.CheckUnary(y, act)
+    else:
+      pre_act = y
+
+    if not isinstance(nn._bias, blocks_std.PassThrough):
+      pre_bias = self.CheckBiasAdd(pre_act, nn._bias)
+    else:
+      pre_bias = pre_act
+
+    if len(nn._matrices) > 1:
+      self.assertEqual('AddN', pre_bias.op.type)
+      pre_bias = pre_bias.op.inputs
+    else:
+      pre_bias = [pre_bias]
+
+    self.assertEqual(len(pre_bias), len(nn._matrices))
+    return [self.CheckLinear(u, m) for u, m in zip(pre_bias, nn._matrices)]
+
+  def testNNWithoutActWithoutBias(self):
+    nn = blocks_std.NN(10, act=None, bias=None)
+    x = tf.placeholder(dtype=tf.float32, shape=[5, 7])
+    y = nn(x)
+    self.assertIs(x, self.CheckNN(y, nn)[0])
+
+  def testNNWithoutBiasWithAct(self):
+    nn = blocks_std.NN(10, act=tf.nn.relu, bias=None)
+    x = tf.placeholder(dtype=tf.float32, shape=[5, 7])
+    y = nn(x)
+    self.assertIs(x, self.CheckNN(y, nn, 'Relu')[0])
+
+  def testNNWithBiasWithoutAct(self):
+    nn = blocks_std.NN(10, bias=blocks_std.Bias(0), act=None)
+    x = tf.placeholder(dtype=tf.float32, shape=[5, 7])
+    y = nn(x)
+    self.assertIs(x, self.CheckNN(y, nn)[0])
+
+  def testNNWithBiasWithAct(self):
+    nn = blocks_std.NN(10, bias=blocks_std.Bias(0), act=tf.square)
+    x = tf.placeholder(dtype=tf.float32, shape=[5, 7])
+    y = nn(x)
+    self.assertIs(x, self.CheckNN(y, nn, 'Square')[0])
+
+  def testNNMultipleInputs(self):
+    nn = blocks_std.NN(10, bias=blocks_std.Bias(0), act=tf.tanh)
+    x = [tf.placeholder(dtype=tf.float32, shape=[5, 7]),
+         tf.placeholder(dtype=tf.float32, shape=[5, 3]),
+         tf.placeholder(dtype=tf.float32, shape=[5, 5])]
+    y = nn(*x)
+    xs = self.CheckNN(y, nn, 'Tanh')
+    self.assertEqual(len(x), len(xs))
+    for u, v in zip(x, xs):
+      self.assertIs(u, v)
+
+  def testConv2DSAME(self):
+    np.random.seed(142536)
+
+    x_shape = [4, 16, 11, 5]
+    f_shape = [4, 3, 5, 6]
+    strides = [1, 2, 2, 1]
+    padding = 'SAME'
+
+    conv = blocks_std.Conv2D(depth=f_shape[-1],
+                             filter_size=f_shape[0:2],
+                             strides=strides[1:3],
+                             padding=padding,
+                             act=None,
+                             bias=None)
+    x_value = np.random.normal(size=x_shape)
+    x = tf.convert_to_tensor(x_value, dtype=tf.float32)
+    y = conv(x)
+
+    with self.test_session():
+      tf.global_variables_initializer().run()
+      f_value = conv._kernel.eval()
+      y_value = y.eval()
+
+    y_expected = _NumpyConv2D(x_value, f_value,
+                              strides=strides, padding=padding)
+    self.assertAllClose(y_expected, y_value)
+
+  def testConv2DValid(self):
+    np.random.seed(253647)
+
+    x_shape = [4, 11, 12, 5]
+    f_shape = [5, 2, 5, 5]
+    strides = [1, 2, 2, 1]
+    padding = 'VALID'
+
+    conv = blocks_std.Conv2D(depth=f_shape[-1],
+                             filter_size=f_shape[0:2],
+                             strides=strides[1:3],
+                             padding=padding,
+                             act=None,
+                             bias=None)
+    x_value = np.random.normal(size=x_shape)
+    x = tf.convert_to_tensor(x_value, dtype=tf.float32)
+    y = conv(x)
+
+    with self.test_session():
+      tf.global_variables_initializer().run()
+      f_value = conv._kernel.eval()
+      y_value = y.eval()
+
+    y_expected = _NumpyConv2D(x_value, f_value,
+                              strides=strides, padding=padding)
+    self.assertAllClose(y_expected, y_value)
+
+  def testConv2DSymmetric(self):
+    np.random.seed(364758)
+
+    x_shape = [4, 10, 12, 6]
+    f_shape = [3, 4, 6, 5]
+    strides = [1, 1, 1, 1]
+    padding = 'SYMMETRIC'
+
+    conv = blocks_std.Conv2D(depth=f_shape[-1],
+                             filter_size=f_shape[0:2],
+                             strides=strides[1:3],
+                             padding=padding,
+                             act=None,
+                             bias=None)
+    x_value = np.random.normal(size=x_shape)
+    x = tf.convert_to_tensor(x_value, dtype=tf.float32)
+    y = conv(x)
+
+    with self.test_session():
+      tf.global_variables_initializer().run()
+      f_value = conv._kernel.eval()
+      y_value = y.eval()
+
+    y_expected = _NumpyConv2D(x_value, f_value,
+                              strides=strides, padding=padding)
+    self.assertAllClose(y_expected, y_value)
+
+  def testConv2DReflect(self):
+    np.random.seed(768798)
+
+    x_shape = [4, 10, 12, 6]
+    f_shape = [3, 4, 6, 5]
+    strides = [1, 2, 2, 1]
+    padding = 'REFLECT'
+
+    conv = blocks_std.Conv2D(depth=f_shape[-1],
+                             filter_size=f_shape[0:2],
+                             strides=strides[1:3],
+                             padding=padding,
+                             act=None,
+                             bias=None)
+    x_value = np.random.normal(size=x_shape)
+    x = tf.convert_to_tensor(x_value, dtype=tf.float32)
+    y = conv(x)
+
+    with self.test_session():
+      tf.global_variables_initializer().run()
+      f_value = conv._kernel.eval()
+      y_value = y.eval()
+
+    y_expected = _NumpyConv2D(x_value, f_value,
+                              strides=strides, padding=padding)
+    self.assertAllClose(y_expected, y_value)
+
+  def testConv2DBias(self):
+    input_shape = [19, 14, 14, 64]
+    filter_shape = [3, 7, 64, 128]
+    strides = [1, 2, 2, 1]
+    output_shape = [19, 6, 4, 128]
+
+    conv = blocks_std.Conv2D(depth=filter_shape[-1],
+                             filter_size=filter_shape[0:2],
+                             strides=strides[1:3],
+                             padding='VALID',
+                             act=None,
+                             bias=blocks_std.Bias(1))
+    x = tf.placeholder(dtype=tf.float32, shape=input_shape)
+
+    y = conv(x)
+    self.CheckBiasAdd(y, conv._bias)
+    self.assertEqual(output_shape, y.get_shape().as_list())
+
+
+if __name__ == '__main__':
+  tf.test.main()

compression/entropy_coder/model/entropy_coder_model.py

+# 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.
+# ==============================================================================
+
+"""Entropy coder model."""
+
+
+class EntropyCoderModel(object):
+  """Entropy coder model."""
+
+  def __init__(self):
+    # Loss used for training the model.
+    self.loss = None
+
+    # Tensorflow op to run to train the model.
+    self.train_op = None
+
+    # Tensor corresponding to the average code length of the input bit field
+    # tensor. The average code length is a number of output bits per input bit.
+    # To get an effective compression, this number should be between 0.0
+    # and 1.0 (1.0 corresponds to no compression).
+    self.average_code_length = None
+
+  def Initialize(self, global_step, optimizer, config_string):
+    raise NotImplementedError()
+
+  def BuildGraph(self, input_codes):
+    """Build the Tensorflow graph corresponding to the entropy coder model.
+
+    Args:
+      input_codes: Tensor of size: batch_size x height x width x bit_depth
+        corresponding to the codes to compress.
+        The input codes are {-1, +1} codes.
+    """
+    # TODO:
+    # - consider switching to {0, 1} codes.
+    # - consider passing an extra tensor which gives for each (b, y, x)
+    #   what is the actual depth (which would allow to use more or less bits
+    #   for each (y, x) location.
+    raise NotImplementedError()
+
+  def GetConfigStringForUnitTest(self):
+    """Returns a default model configuration to be used for unit tests."""
+    return None

compression/entropy_coder/model/model_factory.py

+# 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.
+# ==============================================================================
+
+"""Entropy coder model registrar."""
+
+
+class ModelFactory(object):
+  """Factory of encoder/decoder models."""
+
+  def __init__(self):
+    self._model_dictionary = dict()
+
+  def RegisterModel(self,
+                    entropy_coder_model_name,
+                    entropy_coder_model_factory):
+    self._model_dictionary[entropy_coder_model_name] = (
+        entropy_coder_model_factory)
+
+  def CreateModel(self, model_name):
+    current_model_factory = self._model_dictionary[model_name]
+    return current_model_factory()
+
+  def GetAvailableModels(self):
+    return self._model_dictionary.keys()
+
+
+_model_registry = ModelFactory()
+
+
+def GetModelRegistry():
+  return _model_registry
+
+
+class RegisterEntropyCoderModel(object):
+
+  def __init__(self, model_name):
+    self._model_name = model_name
+
+  def __call__(self, f):
+    _model_registry.RegisterModel(self._model_name, f)
+    return f

compression/entropy_coder/progressive/progressive.py

+# 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.
+# ==============================================================================
+
+"""Code probability model used for entropy coding."""
+
+import json
+
+import tensorflow as tf
+
+from entropy_coder.lib import blocks
+from entropy_coder.model import entropy_coder_model
+from entropy_coder.model import model_factory
+
+# pylint: disable=not-callable
+
+
+class BrnnPredictor(blocks.BlockBase):
+  """BRNN prediction applied on one layer."""
+
+  def __init__(self, code_depth, name=None):
+    super(BrnnPredictor, self).__init__(name)
+
+    with self._BlockScope():
+      hidden_depth = 2 * code_depth
+
+      # What is coming from the previous layer/iteration
+      # is going through a regular Conv2D layer as opposed to the binary codes
+      # of the current layer/iteration which are going through a masked
+      # convolution.
+      self._adaptation0 = blocks.RasterScanConv2D(
+          hidden_depth, [7, 7], [1, 1], 'SAME',
+          strict_order=True,
+          bias=blocks.Bias(0), act=tf.tanh)
+      self._adaptation1 = blocks.Conv2D(
+          hidden_depth, [3, 3], [1, 1], 'SAME',
+          bias=blocks.Bias(0), act=tf.tanh)
+      self._predictor = blocks.CompositionOperator([
+          blocks.LineOperator(
+              blocks.RasterScanConv2DLSTM(
+                  depth=hidden_depth,
+                  filter_size=[1, 3],
+                  hidden_filter_size=[1, 3],
+                  strides=[1, 1],
+                  padding='SAME')),
+          blocks.Conv2D(hidden_depth, [1, 1], [1, 1], 'SAME',
+                        bias=blocks.Bias(0), act=tf.tanh),
+          blocks.Conv2D(code_depth, [1, 1], [1, 1], 'SAME',
+                        bias=blocks.Bias(0), act=tf.tanh)
+      ])
+
+  def _Apply(self, x, s):
+    # Code estimation using both:
+    # - the state from the previous iteration/layer,
+    # - the binary codes that are before in raster scan order.
+    h = tf.concat(values=[self._adaptation0(x), self._adaptation1(s)], axis=3)
+
+    estimated_codes = self._predictor(h)
+
+    return estimated_codes
+
+
+class LayerPrediction(blocks.BlockBase):
+  """Binary code prediction for one layer."""
+
+  def __init__(self, layer_count, code_depth, name=None):
+    super(LayerPrediction, self).__init__(name)
+
+    self._layer_count = layer_count
+
+    # No previous layer.
+    self._layer_state = None
+    self._current_layer = 0
+
+    with self._BlockScope():
+      # Layers used to do the conditional code prediction.
+      self._brnn_predictors = []
+      for _ in xrange(layer_count):
+        self._brnn_predictors.append(BrnnPredictor(code_depth))
+
+      # Layers used to generate the input of the LSTM operating on the
+      # iteration/depth domain.
+      hidden_depth = 2 * code_depth
+      self._state_blocks = []
+      for _ in xrange(layer_count):
+        self._state_blocks.append(blocks.CompositionOperator([
+            blocks.Conv2D(
+                hidden_depth, [3, 3], [1, 1], 'SAME',
+                bias=blocks.Bias(0), act=tf.tanh),
+            blocks.Conv2D(
+                code_depth, [3, 3], [1, 1], 'SAME',
+                bias=blocks.Bias(0), act=tf.tanh)
+        ]))
+
+      # Memory of the RNN is equivalent to the size of 2 layers of binary
+      # codes.
+      hidden_depth = 2 * code_depth
+      self._layer_rnn = blocks.CompositionOperator([
+          blocks.Conv2DLSTM(
+              depth=hidden_depth,
+              filter_size=[1, 1],
+              hidden_filter_size=[1, 1],
+              strides=[1, 1],
+              padding='SAME'),
+          blocks.Conv2D(hidden_depth, [1, 1], [1, 1], 'SAME',
+                        bias=blocks.Bias(0), act=tf.tanh),
+          blocks.Conv2D(code_depth, [1, 1], [1, 1], 'SAME',
+                        bias=blocks.Bias(0), act=tf.tanh)
+      ])
+
+  def _Apply(self, x):
+    assert self._current_layer < self._layer_count
+
+    # Layer state is set to 0 when there is no previous iteration.
+    if self._layer_state is None:
+      self._layer_state = tf.zeros_like(x, dtype=tf.float32)
+
+    # Code estimation using both:
+    # - the state from the previous iteration/layer,
+    # - the binary codes that are before in raster scan order.
+    estimated_codes = self._brnn_predictors[self._current_layer](
+        x, self._layer_state)
+
+    # Compute the updated layer state.
+    h = self._state_blocks[self._current_layer](x)
+    self._layer_state = self._layer_rnn(h)
+    self._current_layer += 1
+
+    return estimated_codes
+
+
+class ProgressiveModel(entropy_coder_model.EntropyCoderModel):
+  """Progressive BRNN entropy coder model."""
+
+  def __init__(self):
+    super(ProgressiveModel, self).__init__()
+
+  def Initialize(self, global_step, optimizer, config_string):
+    if config_string is None:
+      raise ValueError('The progressive model requires a configuration.')
+    config = json.loads(config_string)
+    if 'coded_layer_count' not in config:
+      config['coded_layer_count'] = 0
+
+    self._config = config
+    self._optimizer = optimizer
+    self._global_step = global_step
+
+  def BuildGraph(self, input_codes):
+    """Build the graph corresponding to the progressive BRNN model."""
+    layer_depth = self._config['layer_depth']
+    layer_count = self._config['layer_count']
+
+    code_shape = input_codes.get_shape()
+    code_depth = code_shape[-1].value
+    if self._config['coded_layer_count'] > 0:
+      prefix_depth = self._config['coded_layer_count'] * layer_depth
+      if code_depth < prefix_depth:
+        raise ValueError('Invalid prefix depth: {} VS {}'.format(
+            prefix_depth, code_depth))
+      input_codes = input_codes[:, :, :, :prefix_depth]
+
+    code_shape = input_codes.get_shape()
+    code_depth = code_shape[-1].value
+    if code_depth % layer_depth != 0:
+      raise ValueError(
+          'Code depth must be a multiple of the layer depth: {} vs {}'.format(
+              code_depth, layer_depth))
+    code_layer_count = code_depth // layer_depth
+    if code_layer_count > layer_count:
+      raise ValueError('Input codes have too many layers: {}, max={}'.format(
+          code_layer_count, layer_count))
+
+    # Block used to estimate binary codes.
+    layer_prediction = LayerPrediction(layer_count, layer_depth)
+
+    # Block used to compute code lengths.
+    code_length_block = blocks.CodeLength()
+
+    # Loop over all the layers.
+    code_length = []
+    code_layers = tf.split(
+        value=input_codes, num_or_size_splits=code_layer_count, axis=3)
+    for k in xrange(code_layer_count):
+      x = code_layers[k]
+      predicted_x = layer_prediction(x)
+      # Saturate the prediction to avoid infinite code length.
+      epsilon = 0.001
+      predicted_x = tf.clip_by_value(
+          predicted_x, -1 + epsilon, +1 - epsilon)
+      code_length.append(code_length_block(
+          blocks.ConvertSignCodeToZeroOneCode(x),
+          blocks.ConvertSignCodeToZeroOneCode(predicted_x)))
+      tf.contrib.deprecated.scalar_summary('code_length_layer_{:02d}'.format(k),
+                                           code_length[-1])
+    code_length = tf.stack(code_length)
+    self.loss = tf.reduce_mean(code_length)
+    tf.contrib.deprecated.scalar_summary('loss', self.loss)
+
+    # Loop over all the remaining layers just to make sure they are
+    # instantiated. Otherwise, loading model params could fail.
+    dummy_x = tf.zeros_like(code_layers[0])
+    for _ in xrange(layer_count - code_layer_count):
+      dummy_predicted_x = layer_prediction(dummy_x)
+
+    # Average bitrate over total_line_count.
+    self.average_code_length = tf.reduce_mean(code_length)
+
+    if self._optimizer:
+      optim_op = self._optimizer.minimize(self.loss,
+                                          global_step=self._global_step)
+      block_updates = blocks.CreateBlockUpdates()
+      if block_updates:
+        with tf.get_default_graph().control_dependencies([optim_op]):
+          self.train_op = tf.group(*block_updates)
+      else:
+        self.train_op = optim_op
+    else:
+      self.train_op = None
+
+  def GetConfigStringForUnitTest(self):
+    s = '{\n'
+    s += '"layer_depth": 1,\n'
+    s += '"layer_count": 8\n'
+    s += '}\n'
+    return s
+
+
+@model_factory.RegisterEntropyCoderModel('progressive')
+def CreateProgressiveModel():
+  return ProgressiveModel()