resnet.py

# Copyright 2020 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 of Residual Networks."""

# Import libraries
import tensorflow as tf
from official.modeling import tf_utils
from official.vision.beta.modeling.backbones import factory
from official.vision.beta.modeling.layers import nn_blocks
from official.vision.beta.modeling.layers import nn_layers

layers = tf.keras.layers

# Specifications for different ResNet variants.
# Each entry specifies block configurations of the particular ResNet variant.
# Each element in the block configuration is in the following format:
# (block_fn, num_filters, block_repeats)
RESNET_SPECS = {
    18: [
        ('residual', 64, 2),
        ('residual', 128, 2),
        ('residual', 256, 2),
        ('residual', 512, 2),
    ],
    34: [
        ('residual', 64, 3),
        ('residual', 128, 4),
        ('residual', 256, 6),
        ('residual', 512, 3),
    ],
    50: [
        ('bottleneck', 64, 3),
        ('bottleneck', 128, 4),
        ('bottleneck', 256, 6),
        ('bottleneck', 512, 3),
    ],
    101: [
        ('bottleneck', 64, 3),
        ('bottleneck', 128, 4),
        ('bottleneck', 256, 23),
        ('bottleneck', 512, 3),
    ],
    152: [
        ('bottleneck', 64, 3),
        ('bottleneck', 128, 8),
        ('bottleneck', 256, 36),
        ('bottleneck', 512, 3),
    ],
    200: [
        ('bottleneck', 64, 3),
        ('bottleneck', 128, 24),
        ('bottleneck', 256, 36),
        ('bottleneck', 512, 3),
    ],
    270: [
        ('bottleneck', 64, 4),
        ('bottleneck', 128, 29),
        ('bottleneck', 256, 53),
        ('bottleneck', 512, 4),
    ],
    350: [
        ('bottleneck', 64, 4),
        ('bottleneck', 128, 36),
        ('bottleneck', 256, 72),
        ('bottleneck', 512, 4),
    ],
    420: [
        ('bottleneck', 64, 4),
        ('bottleneck', 128, 44),
        ('bottleneck', 256, 87),
        ('bottleneck', 512, 4),
    ],
}


@tf.keras.utils.register_keras_serializable(package='Vision')
class ResNet(tf.keras.Model):
  """Creates a ResNet family model.

  This implements the Deep Residual Network from:
    Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun.
    Deep Residual Learning for Image Recognition.
    (https://arxiv.org/pdf/1512.03385)
  """

  def __init__(self,
               model_id,
               input_specs=layers.InputSpec(shape=[None, None, None, 3]),
               depth_multiplier=1.0,
               stem_type='v0',
               resnetd_shortcut=False,
               replace_stem_max_pool=False,
               se_ratio=None,
               init_stochastic_depth_rate=0.0,
               activation='relu',
               use_sync_bn=False,
               norm_momentum=0.99,
               norm_epsilon=0.001,
               kernel_initializer='VarianceScaling',
               kernel_regularizer=None,
               bias_regularizer=None,
               **kwargs):
    """Initializes a ResNet model.

    Args:
      model_id: An `int` of the depth of ResNet backbone model.
      input_specs: A `tf.keras.layers.InputSpec` of the input tensor.
      depth_multiplier: A `float` of the depth multiplier to uniformaly scale up
        all layers in channel size in ResNet.
      stem_type: A `str` of stem type of ResNet. Default to `v0`. If set to
        `v1`, use ResNet-D type stem (https://arxiv.org/abs/1812.01187).
      resnetd_shortcut: A `bool` of whether to use ResNet-D shortcut in
        downsampling blocks.
      replace_stem_max_pool: A `bool` of whether to replace the max pool in stem
        with a stride-2 conv,
      se_ratio: A `float` or None. Ratio of the Squeeze-and-Excitation layer.
      init_stochastic_depth_rate: A `float` of initial stochastic depth rate.
      activation: A `str` name of the activation function.
      use_sync_bn: If True, use synchronized batch normalization.
      norm_momentum: A `float` of normalization momentum for the moving average.
      norm_epsilon: A small `float` added to variance to avoid dividing by zero.
      kernel_initializer: A str for kernel initializer of convolutional layers.
      kernel_regularizer: A `tf.keras.regularizers.Regularizer` object for
        Conv2D. Default to None.
      bias_regularizer: A `tf.keras.regularizers.Regularizer` object for Conv2D.
        Default to None.
      **kwargs: Additional keyword arguments to be passed.
    """
    self._model_id = model_id
    self._input_specs = input_specs
    self._depth_multiplier = depth_multiplier
    self._stem_type = stem_type
    self._resnetd_shortcut = resnetd_shortcut
    self._replace_stem_max_pool = replace_stem_max_pool
    self._se_ratio = se_ratio
    self._init_stochastic_depth_rate = init_stochastic_depth_rate
    self._use_sync_bn = use_sync_bn
    self._activation = activation
    self._norm_momentum = norm_momentum
    self._norm_epsilon = norm_epsilon
    if use_sync_bn:
      self._norm = layers.experimental.SyncBatchNormalization
    else:
      self._norm = layers.BatchNormalization
    self._kernel_initializer = kernel_initializer
    self._kernel_regularizer = kernel_regularizer
    self._bias_regularizer = bias_regularizer

    if tf.keras.backend.image_data_format() == 'channels_last':
      bn_axis = -1
    else:
      bn_axis = 1

    # Build ResNet.
    inputs = tf.keras.Input(shape=input_specs.shape[1:])

    if stem_type == 'v0':
      x = layers.Conv2D(
          filters=int(64 * self._depth_multiplier),
          kernel_size=7,
          strides=2,
          use_bias=False,
          padding='same',
          kernel_initializer=self._kernel_initializer,
          kernel_regularizer=self._kernel_regularizer,
          bias_regularizer=self._bias_regularizer)(
              inputs)
      x = self._norm(
          axis=bn_axis, momentum=norm_momentum, epsilon=norm_epsilon)(
              x)
      x = tf_utils.get_activation(activation)(x)
    elif stem_type == 'v1':
      x = layers.Conv2D(
          filters=int(32 * self._depth_multiplier),
          kernel_size=3,
          strides=2,
          use_bias=False,
          padding='same',
          kernel_initializer=self._kernel_initializer,
          kernel_regularizer=self._kernel_regularizer,
          bias_regularizer=self._bias_regularizer)(
              inputs)
      x = self._norm(
          axis=bn_axis, momentum=norm_momentum, epsilon=norm_epsilon)(
              x)
      x = tf_utils.get_activation(activation)(x)
      x = layers.Conv2D(
          filters=int(32 * self._depth_multiplier),
          kernel_size=3,
          strides=1,
          use_bias=False,
          padding='same',
          kernel_initializer=self._kernel_initializer,
          kernel_regularizer=self._kernel_regularizer,
          bias_regularizer=self._bias_regularizer)(
              x)
      x = self._norm(
          axis=bn_axis, momentum=norm_momentum, epsilon=norm_epsilon)(
              x)
      x = tf_utils.get_activation(activation)(x)
      x = layers.Conv2D(
          filters=int(64 * self._depth_multiplier),
          kernel_size=3,
          strides=1,
          use_bias=False,
          padding='same',
          kernel_initializer=self._kernel_initializer,
          kernel_regularizer=self._kernel_regularizer,
          bias_regularizer=self._bias_regularizer)(
              x)
      x = self._norm(
          axis=bn_axis, momentum=norm_momentum, epsilon=norm_epsilon)(
              x)
      x = tf_utils.get_activation(activation)(x)
    else:
      raise ValueError('Stem type {} not supported.'.format(stem_type))

    if replace_stem_max_pool:
      x = layers.Conv2D(
          filters=int(64 * self._depth_multiplier),
          kernel_size=3,
          strides=2,
          use_bias=False,
          padding='same',
          kernel_initializer=self._kernel_initializer,
          kernel_regularizer=self._kernel_regularizer,
          bias_regularizer=self._bias_regularizer)(
              x)
      x = self._norm(
          axis=bn_axis, momentum=norm_momentum, epsilon=norm_epsilon)(
              x)
      x = tf_utils.get_activation(activation)(x)
    else:
      x = layers.MaxPool2D(pool_size=3, strides=2, padding='same')(x)

    endpoints = {}
    for i, spec in enumerate(RESNET_SPECS[model_id]):
      if spec[0] == 'residual':
        block_fn = nn_blocks.ResidualBlock
      elif spec[0] == 'bottleneck':
        block_fn = nn_blocks.BottleneckBlock
      else:
        raise ValueError('Block fn `{}` is not supported.'.format(spec[0]))
      x = self._block_group(
          inputs=x,
          filters=int(spec[1] * self._depth_multiplier),
          strides=(1 if i == 0 else 2),
          block_fn=block_fn,
          block_repeats=spec[2],
          stochastic_depth_drop_rate=nn_layers.get_stochastic_depth_rate(
              self._init_stochastic_depth_rate, i + 2, 5),
          name='block_group_l{}'.format(i + 2))
      endpoints[str(i + 2)] = x

    self._output_specs = {l: endpoints[l].get_shape() for l in endpoints}

    super(ResNet, self).__init__(inputs=inputs, outputs=endpoints, **kwargs)

  def _block_group(self,
                   inputs,
                   filters,
                   strides,
                   block_fn,
                   block_repeats=1,
                   stochastic_depth_drop_rate=0.0,
                   name='block_group'):
    """Creates one group of blocks for the ResNet model.

    Args:
      inputs: A `tf.Tensor` of size `[batch, channels, height, width]`.
      filters: An `int` number of filters for the first convolution of the
        layer.
      strides: An `int` stride to use for the first convolution of the layer.
        If greater than 1, this layer will downsample the input.
      block_fn: The type of block group. Either `nn_blocks.ResidualBlock` or
        `nn_blocks.BottleneckBlock`.
      block_repeats: An `int` number of blocks contained in the layer.
      stochastic_depth_drop_rate: A `float` of drop rate of the current block
        group.
      name: A `str` name for the block.

    Returns:
      The output `tf.Tensor` of the block layer.
    """
    x = block_fn(
        filters=filters,
        strides=strides,
        use_projection=True,
        stochastic_depth_drop_rate=stochastic_depth_drop_rate,
        se_ratio=self._se_ratio,
        resnetd_shortcut=self._resnetd_shortcut,
        kernel_initializer=self._kernel_initializer,
        kernel_regularizer=self._kernel_regularizer,
        bias_regularizer=self._bias_regularizer,
        activation=self._activation,
        use_sync_bn=self._use_sync_bn,
        norm_momentum=self._norm_momentum,
        norm_epsilon=self._norm_epsilon)(
            inputs)

    for _ in range(1, block_repeats):
      x = block_fn(
          filters=filters,
          strides=1,
          use_projection=False,
          stochastic_depth_drop_rate=stochastic_depth_drop_rate,
          se_ratio=self._se_ratio,
          resnetd_shortcut=self._resnetd_shortcut,
          kernel_initializer=self._kernel_initializer,
          kernel_regularizer=self._kernel_regularizer,
          bias_regularizer=self._bias_regularizer,
          activation=self._activation,
          use_sync_bn=self._use_sync_bn,
          norm_momentum=self._norm_momentum,
          norm_epsilon=self._norm_epsilon)(
              x)

    return tf.identity(x, name=name)

  def get_config(self):
    config_dict = {
        'model_id': self._model_id,
        'depth_multiplier': self._depth_multiplier,
        'stem_type': self._stem_type,
        'resnetd_shortcut': self._resnetd_shortcut,
        'replace_stem_max_pool': self._replace_stem_max_pool,
        'activation': self._activation,
        'se_ratio': self._se_ratio,
        'init_stochastic_depth_rate': self._init_stochastic_depth_rate,
        'use_sync_bn': self._use_sync_bn,
        'norm_momentum': self._norm_momentum,
        'norm_epsilon': self._norm_epsilon,
        'kernel_initializer': self._kernel_initializer,
        'kernel_regularizer': self._kernel_regularizer,
        'bias_regularizer': self._bias_regularizer,
    }
    return config_dict

  @classmethod
  def from_config(cls, config, custom_objects=None):
    return cls(**config)

  @property
  def output_specs(self):
    """A dict of {level: TensorShape} pairs for the model output."""
    return self._output_specs


@factory.register_backbone_builder('resnet')
def build_resnet(
    input_specs: tf.keras.layers.InputSpec,
    model_config,
    l2_regularizer: tf.keras.regularizers.Regularizer = None) -> tf.keras.Model:
  """Builds ResNet backbone from a config."""
  backbone_type = model_config.backbone.type
  backbone_cfg = model_config.backbone.get()
  norm_activation_config = model_config.norm_activation
  assert backbone_type == 'resnet', (f'Inconsistent backbone type '
                                     f'{backbone_type}')

  return ResNet(
      model_id=backbone_cfg.model_id,
      input_specs=input_specs,
      depth_multiplier=backbone_cfg.depth_multiplier,
      stem_type=backbone_cfg.stem_type,
      resnetd_shortcut=backbone_cfg.resnetd_shortcut,
      replace_stem_max_pool=backbone_cfg.replace_stem_max_pool,
      se_ratio=backbone_cfg.se_ratio,
      init_stochastic_depth_rate=backbone_cfg.stochastic_depth_drop_rate,
      activation=norm_activation_config.activation,
      use_sync_bn=norm_activation_config.use_sync_bn,
      norm_momentum=norm_activation_config.norm_momentum,
      norm_epsilon=norm_activation_config.norm_epsilon,
      kernel_regularizer=l2_regularizer)