resnet_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.
# ==============================================================================
"""Contains definitions for the preactivation form of Residual Networks.

Residual networks (ResNets) were originally proposed in:
[1] Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun
    Deep Residual Learning for Image Recognition. arXiv:1512.03385

The full preactivation 'v2' ResNet variant implemented in this module was
introduced by:
[2] Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun
    Identity Mappings in Deep Residual Networks. arXiv: 1603.05027

The key difference of the full preactivation 'v2' variant compared to the
'v1' variant in [1] is the use of batch normalization before every weight layer
rather than after.
"""

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import tensorflow as tf

_BATCH_NORM_DECAY = 0.997
_BATCH_NORM_EPSILON = 1e-5


def batch_norm_relu(inputs, training, data_format):
  """Performs a batch normalization followed by a ReLU."""
  # We set fused=True for a significant performance boost. See
  # https://www.tensorflow.org/performance/performance_guide#common_fused_ops
  inputs = tf.layers.batch_normalization(
      inputs=inputs, axis=1 if data_format == 'channels_first' else 3,
      momentum=_BATCH_NORM_DECAY, epsilon=_BATCH_NORM_EPSILON, center=True,
      scale=True, training=training, fused=True)
  inputs = tf.nn.relu(inputs)
  return inputs


def fixed_padding(inputs, kernel_size, data_format):
  """Pads the input along the spatial dimensions independently of input size.

  Args:
    inputs: A tensor of size [batch, channels, height_in, width_in] or
      [batch, height_in, width_in, channels] depending on data_format.
    kernel_size: The kernel to be used in the conv2d or max_pool2d operation.
                 Should be a positive integer.
    data_format: The input format ('channels_last' or 'channels_first').

  Returns:
    A tensor with the same format as the input with the data either intact
    (if kernel_size == 1) or padded (if kernel_size > 1).
  """
  pad_total = kernel_size - 1
  pad_beg = pad_total // 2
  pad_end = pad_total - pad_beg

  if data_format == 'channels_first':
    padded_inputs = tf.pad(inputs, [[0, 0], [0, 0],
                                    [pad_beg, pad_end], [pad_beg, pad_end]])
  else:
    padded_inputs = tf.pad(inputs, [[0, 0], [pad_beg, pad_end],
                                    [pad_beg, pad_end], [0, 0]])
  return padded_inputs


def conv2d_fixed_padding(inputs, filters, kernel_size, strides, data_format):
  """Strided 2-D convolution with explicit padding."""
  # The padding is consistent and is based only on `kernel_size`, not on the
  # dimensions of `inputs` (as opposed to using `tf.layers.conv2d` alone).
  if strides > 1:
    inputs = fixed_padding(inputs, kernel_size, data_format)

  return tf.layers.conv2d(
      inputs=inputs, filters=filters, kernel_size=kernel_size, strides=strides,
      padding=('SAME' if strides == 1 else 'VALID'), use_bias=False,
      kernel_initializer=tf.variance_scaling_initializer(),
      data_format=data_format)


def building_block(inputs, filters, training, projection_shortcut, strides,
                   data_format):
  """Standard building block for residual networks with BN before convolutions.

  Args:
    inputs: A tensor of size [batch, channels, height_in, width_in] or
      [batch, height_in, width_in, channels] depending on data_format.
    filters: The number of filters for the convolutions.
    training: A Boolean for whether the model is in training or inference
      mode. Needed for batch normalization.
    projection_shortcut: The function to use for projection shortcuts
      (typically a 1x1 convolution when downsampling the input).
    strides: The block's stride. If greater than 1, this block will ultimately
      downsample the input.
    data_format: The input format ('channels_last' or 'channels_first').

  Returns:
    The output tensor of the block.
  """
  shortcut = inputs
  inputs = batch_norm_relu(inputs, training, data_format)

  # The projection shortcut should come after the first batch norm and ReLU
  # since it performs a 1x1 convolution.
  if projection_shortcut is not None:
    shortcut = projection_shortcut(inputs)

  inputs = conv2d_fixed_padding(
      inputs=inputs, filters=filters, kernel_size=3, strides=strides,
      data_format=data_format)

  inputs = batch_norm_relu(inputs, training, data_format)
  inputs = conv2d_fixed_padding(
      inputs=inputs, filters=filters, kernel_size=3, strides=1,
      data_format=data_format)

  return inputs + shortcut


def bottleneck_block(inputs, filters, training, projection_shortcut,
                     strides, data_format):
  """Bottleneck block variant for residual networks with BN before convolutions.

  Args:
    inputs: A tensor of size [batch, channels, height_in, width_in] or
      [batch, height_in, width_in, channels] depending on data_format.
    filters: The number of filters for the first two convolutions. Note
      that the third and final convolution will use 4 times as many filters.
    training: A Boolean for whether the model is in training or inference
      mode. Needed for batch normalization.
    projection_shortcut: The function to use for projection shortcuts
      (typically a 1x1 convolution when downsampling the input).
    strides: The block's stride. If greater than 1, this block will ultimately
      downsample the input.
    data_format: The input format ('channels_last' or 'channels_first').

  Returns:
    The output tensor of the block.
  """
  shortcut = inputs
  inputs = batch_norm_relu(inputs, training, data_format)

  # The projection shortcut should come after the first batch norm and ReLU
  # since it performs a 1x1 convolution.
  if projection_shortcut is not None:
    shortcut = projection_shortcut(inputs)

  inputs = conv2d_fixed_padding(
      inputs=inputs, filters=filters, kernel_size=1, strides=1,
      data_format=data_format)

  inputs = batch_norm_relu(inputs, training, data_format)
  inputs = conv2d_fixed_padding(
      inputs=inputs, filters=filters, kernel_size=3, strides=strides,
      data_format=data_format)

  inputs = batch_norm_relu(inputs, training, data_format)
  inputs = conv2d_fixed_padding(
      inputs=inputs, filters=4 * filters, kernel_size=1, strides=1,
      data_format=data_format)

  return inputs + shortcut


def block_layer(inputs, filters, block_fn, blocks, strides, training, name,
                data_format):
  """Creates one layer of blocks for the ResNet model.

  Args:
    inputs: A tensor of size [batch, channels, height_in, width_in] or
      [batch, height_in, width_in, channels] depending on data_format.
    filters: The number of filters for the first convolution of the layer.
    block_fn: The block to use within the model, either `building_block` or
      `bottleneck_block`.
    blocks: The number of blocks contained in the layer.
    strides: The stride to use for the first convolution of the layer. If
      greater than 1, this layer will ultimately downsample the input.
    training: Either True or False, whether we are currently training the
      model. Needed for batch norm.
    name: A string name for the tensor output of the block layer.
    data_format: The input format ('channels_last' or 'channels_first').

  Returns:
    The output tensor of the block layer.
  """
  # Bottleneck blocks end with 4x the number of filters as they start with
  filters_out = 4 * filters if block_fn is bottleneck_block else filters

  def projection_shortcut(inputs):
    return conv2d_fixed_padding(
        inputs=inputs, filters=filters_out, kernel_size=1, strides=strides,
        data_format=data_format)

  # Only the first block per block_layer uses projection_shortcut and strides
  inputs = block_fn(inputs, filters, training, projection_shortcut, strides,
                    data_format)

  for _ in range(1, blocks):
    inputs = block_fn(inputs, filters, training, None, 1, data_format)

  return tf.identity(inputs, name)


class Model(object):
  """Base class for building the Resnet v2 Model.
  """

  def __init__(self, resnet_size, num_classes, num_filters, kernel_size,
               conv_stride, first_pool_size, first_pool_stride,
               second_pool_size, second_pool_stride, block_fn, block_sizes,
               block_strides, final_size, data_format=None):
    """Creates a model for classifying an image.

    Args:
      resnet_size: A single integer for the size of the ResNet model.
      num_classes: The number of classes used as labels.
      num_filters: The number of filters to use for the first block layer
        of the model. This number is then doubled for each subsequent block
        layer.
      kernel_size: The kernel size to use for convolution.
      conv_stride: stride size for the initial convolutional layer
      first_pool_size: Pool size to be used for the first pooling layer.
        If none, the first pooling layer is skipped.
      first_pool_stride: stride size for the first pooling layer. Not used
        if first_pool_size is None.
      second_pool_size: Pool size to be used for the second pooling layer.
      second_pool_stride: stride size for the final pooling layer
      block_fn: Which block layer function should be used? Pass in one of
        the two functions defined above: building_block or bottleneck_block
      block_sizes: A list containing n values, where n is the number of sets of
        block layers desired. Each value should be the number of blocks in the
        i-th set.
      block_strides: List of integers representing the desired stride size for
        each of the sets of block layers. Should be same length as block_sizes.
      final_size: The expected size of the model after the second pooling.
      data_format: Input format ('channels_last', 'channels_first', or None).
        If set to None, the format is dependent on whether a GPU is available.
    """
    self.resnet_size = resnet_size

    if not data_format:
      data_format = (
          'channels_first' if tf.test.is_built_with_cuda() else 'channels_last')

    self.data_format = data_format
    self.num_classes = num_classes
    self.num_filters = num_filters
    self.kernel_size = kernel_size
    self.conv_stride = conv_stride
    self.first_pool_size = first_pool_size
    self.first_pool_stride = first_pool_stride
    self.second_pool_size = second_pool_size
    self.second_pool_stride = second_pool_stride
    self.block_fn = block_fn
    self.block_sizes = block_sizes
    self.block_strides = block_strides
    self.final_size = final_size

  def __call__(self, inputs, training):
    """Add operations to classify a batch of input images.

    Args:
      inputs: A Tensor representing a batch of input images.
      training: A boolean. Set to True to add operations required only when
        training the classifier.

    Returns:
      A logits Tensor with shape [<batch_size>, self.num_classes].
    """

    if self.data_format == 'channels_first':
      # Convert the inputs from channels_last (NHWC) to channels_first (NCHW).
      # This provides a large performance boost on GPU. See
      # https://www.tensorflow.org/performance/performance_guide#data_formats
      inputs = tf.transpose(inputs, [0, 3, 1, 2])

    inputs = conv2d_fixed_padding(
        inputs=inputs, filters=self.num_filters, kernel_size=self.kernel_size,
        strides=self.conv_stride, data_format=self.data_format)
    inputs = tf.identity(inputs, 'initial_conv')

    if self.first_pool_size:
      inputs = tf.layers.max_pooling2d(
          inputs=inputs, pool_size=self.first_pool_size,
          strides=self.first_pool_stride, padding='SAME',
          data_format=self.data_format)
      inputs = tf.identity(inputs, 'initial_max_pool')

    for i, num_blocks in enumerate(self.block_sizes):
      num_filters = self.num_filters * (2**i)
      inputs = block_layer(
          inputs=inputs, filters=num_filters, block_fn=self.block_fn,
          blocks=num_blocks, strides=self.block_strides[i],
          training=training, name='block_layer{}'.format(i + 1),
          data_format=self.data_format)

    inputs = batch_norm_relu(inputs, training, self.data_format)
    inputs = tf.layers.average_pooling2d(
        inputs=inputs, pool_size=self.second_pool_size,
        strides=self.second_pool_stride, padding='VALID',
        data_format=self.data_format)
    inputs = tf.identity(inputs, 'final_avg_pool')

    inputs = tf.reshape(inputs, [-1, self.final_size])
    inputs = tf.layers.dense(inputs=inputs, units=self.num_classes)
    inputs = tf.identity(inputs, 'final_dense')
    return inputs