nasnet.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 the definition for the NASNet classification networks.

Paper: https://arxiv.org/abs/1707.07012
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import copy
import tensorflow as tf

from nets.nasnet import nasnet_utils

arg_scope = tf.contrib.framework.arg_scope
slim = tf.contrib.slim


# Notes for training NASNet Cifar Model
# -------------------------------------
# batch_size: 32
# learning rate: 0.025
# cosine (single period) learning rate decay
# auxiliary head loss weighting: 0.4
# clip global norm of all gradients by 5
def cifar_config():
  return tf.contrib.training.HParams(
      stem_multiplier=3.0,
      drop_path_keep_prob=0.6,
      num_cells=18,
      use_aux_head=1,
      num_conv_filters=32,
      dense_dropout_keep_prob=1.0,
      filter_scaling_rate=2.0,
      num_reduction_layers=2,
      data_format='NHWC',
      skip_reduction_layer_input=0,
      # 600 epochs with a batch size of 32
      # This is used for the drop path probabilities since it needs to increase
      # the drop out probability over the course of training.
      total_training_steps=937500,
  )


# Notes for training large NASNet model on ImageNet
# -------------------------------------
# batch size (per replica): 16
# learning rate: 0.015 * 100
# learning rate decay factor: 0.97
# num epochs per decay: 2.4
# sync sgd with 100 replicas
# auxiliary head loss weighting: 0.4
# label smoothing: 0.1
# clip global norm of all gradients by 10
def large_imagenet_config():
  return tf.contrib.training.HParams(
      stem_multiplier=3.0,
      dense_dropout_keep_prob=0.5,
      num_cells=18,
      filter_scaling_rate=2.0,
      num_conv_filters=168,
      drop_path_keep_prob=0.7,
      use_aux_head=1,
      num_reduction_layers=2,
      data_format='NHWC',
      skip_reduction_layer_input=1,
      total_training_steps=250000,
  )


# Notes for training the mobile NASNet ImageNet model
# -------------------------------------
# batch size (per replica): 32
# learning rate: 0.04 * 50
# learning rate scaling factor: 0.97
# num epochs per decay: 2.4
# sync sgd with 50 replicas
# auxiliary head weighting: 0.4
# label smoothing: 0.1
# clip global norm of all gradients by 10
def mobile_imagenet_config():
  return tf.contrib.training.HParams(
      stem_multiplier=1.0,
      dense_dropout_keep_prob=0.5,
      num_cells=12,
      filter_scaling_rate=2.0,
      drop_path_keep_prob=1.0,
      num_conv_filters=44,
      use_aux_head=1,
      num_reduction_layers=2,
      data_format='NHWC',
      skip_reduction_layer_input=0,
      total_training_steps=250000,
  )


def _update_hparams(hparams, is_training):
  """Update hparams for given is_training option."""
  if not is_training:
    hparams.set_hparam('drop_path_keep_prob', 1.0)


def nasnet_cifar_arg_scope(weight_decay=5e-4,
                           batch_norm_decay=0.9,
                           batch_norm_epsilon=1e-5):
  """Defines the default arg scope for the NASNet-A Cifar model.

  Args:
    weight_decay: The weight decay to use for regularizing the model.
    batch_norm_decay: Decay for batch norm moving average.
    batch_norm_epsilon: Small float added to variance to avoid dividing by zero
      in batch norm.

  Returns:
    An `arg_scope` to use for the NASNet Cifar Model.
  """
  batch_norm_params = {
      # Decay for the moving averages.
      'decay': batch_norm_decay,
      # epsilon to prevent 0s in variance.
      'epsilon': batch_norm_epsilon,
      'scale': True,
      'fused': True,
  }
  weights_regularizer = tf.contrib.layers.l2_regularizer(weight_decay)
  weights_initializer = tf.contrib.layers.variance_scaling_initializer(
      mode='FAN_OUT')
  with arg_scope([slim.fully_connected, slim.conv2d, slim.separable_conv2d],
                 weights_regularizer=weights_regularizer,
                 weights_initializer=weights_initializer):
    with arg_scope([slim.fully_connected],
                   activation_fn=None, scope='FC'):
      with arg_scope([slim.conv2d, slim.separable_conv2d],
                     activation_fn=None, biases_initializer=None):
        with arg_scope([slim.batch_norm], **batch_norm_params) as sc:
          return sc


def nasnet_mobile_arg_scope(weight_decay=4e-5,
                            batch_norm_decay=0.9997,
                            batch_norm_epsilon=1e-3):
  """Defines the default arg scope for the NASNet-A Mobile ImageNet model.

  Args:
    weight_decay: The weight decay to use for regularizing the model.
    batch_norm_decay: Decay for batch norm moving average.
    batch_norm_epsilon: Small float added to variance to avoid dividing by zero
      in batch norm.

  Returns:
    An `arg_scope` to use for the NASNet Mobile Model.
  """
  batch_norm_params = {
      # Decay for the moving averages.
      'decay': batch_norm_decay,
      # epsilon to prevent 0s in variance.
      'epsilon': batch_norm_epsilon,
      'scale': True,
      'fused': True,
  }
  weights_regularizer = tf.contrib.layers.l2_regularizer(weight_decay)
  weights_initializer = tf.contrib.layers.variance_scaling_initializer(
      mode='FAN_OUT')
  with arg_scope([slim.fully_connected, slim.conv2d, slim.separable_conv2d],
                 weights_regularizer=weights_regularizer,
                 weights_initializer=weights_initializer):
    with arg_scope([slim.fully_connected],
                   activation_fn=None, scope='FC'):
      with arg_scope([slim.conv2d, slim.separable_conv2d],
                     activation_fn=None, biases_initializer=None):
        with arg_scope([slim.batch_norm], **batch_norm_params) as sc:
          return sc


def nasnet_large_arg_scope(weight_decay=5e-5,
                           batch_norm_decay=0.9997,
                           batch_norm_epsilon=1e-3):
  """Defines the default arg scope for the NASNet-A Large ImageNet model.

  Args:
    weight_decay: The weight decay to use for regularizing the model.
    batch_norm_decay: Decay for batch norm moving average.
    batch_norm_epsilon: Small float added to variance to avoid dividing by zero
      in batch norm.

  Returns:
    An `arg_scope` to use for the NASNet Large Model.
  """
  batch_norm_params = {
      # Decay for the moving averages.
      'decay': batch_norm_decay,
      # epsilon to prevent 0s in variance.
      'epsilon': batch_norm_epsilon,
      'scale': True,
      'fused': True,
  }
  weights_regularizer = tf.contrib.layers.l2_regularizer(weight_decay)
  weights_initializer = tf.contrib.layers.variance_scaling_initializer(
      mode='FAN_OUT')
  with arg_scope([slim.fully_connected, slim.conv2d, slim.separable_conv2d],
                 weights_regularizer=weights_regularizer,
                 weights_initializer=weights_initializer):
    with arg_scope([slim.fully_connected],
                   activation_fn=None, scope='FC'):
      with arg_scope([slim.conv2d, slim.separable_conv2d],
                     activation_fn=None, biases_initializer=None):
        with arg_scope([slim.batch_norm], **batch_norm_params) as sc:
          return sc


def _build_aux_head(net, end_points, num_classes, hparams, scope):
  """Auxiliary head used for all models across all datasets."""
  with tf.variable_scope(scope):
    aux_logits = tf.identity(net)
    with tf.variable_scope('aux_logits'):
      aux_logits = slim.avg_pool2d(
          aux_logits, [5, 5], stride=3, padding='VALID')
      aux_logits = slim.conv2d(aux_logits, 128, [1, 1], scope='proj')
      aux_logits = slim.batch_norm(aux_logits, scope='aux_bn0')
      aux_logits = tf.nn.relu(aux_logits)
      # Shape of feature map before the final layer.
      shape = aux_logits.shape
      if hparams.data_format == 'NHWC':
        shape = shape[1:3]
      else:
        shape = shape[2:4]
      aux_logits = slim.conv2d(aux_logits, 768, shape, padding='VALID')
      aux_logits = slim.batch_norm(aux_logits, scope='aux_bn1')
      aux_logits = tf.nn.relu(aux_logits)
      aux_logits = tf.contrib.layers.flatten(aux_logits)
      aux_logits = slim.fully_connected(aux_logits, num_classes)
      end_points['AuxLogits'] = aux_logits


def _imagenet_stem(inputs, hparams, stem_cell, current_step=None):
  """Stem used for models trained on ImageNet."""
  num_stem_cells = 2

  # 149 x 149 x 32
  num_stem_filters = int(32 * hparams.stem_multiplier)
  net = slim.conv2d(
      inputs, num_stem_filters, [3, 3], stride=2, scope='conv0',
      padding='VALID')
  net = slim.batch_norm(net, scope='conv0_bn')

  # Run the reduction cells
  cell_outputs = [None, net]
  filter_scaling = 1.0 / (hparams.filter_scaling_rate**num_stem_cells)
  for cell_num in range(num_stem_cells):
    net = stem_cell(
        net,
        scope='cell_stem_{}'.format(cell_num),
        filter_scaling=filter_scaling,
        stride=2,
        prev_layer=cell_outputs[-2],
        cell_num=cell_num,
        current_step=current_step)
    cell_outputs.append(net)
    filter_scaling *= hparams.filter_scaling_rate
  return net, cell_outputs


def _cifar_stem(inputs, hparams):
  """Stem used for models trained on Cifar."""
  num_stem_filters = int(hparams.num_conv_filters * hparams.stem_multiplier)
  net = slim.conv2d(
      inputs,
      num_stem_filters,
      3,
      scope='l1_stem_3x3')
  net = slim.batch_norm(net, scope='l1_stem_bn')
  return net, [None, net]


def build_nasnet_cifar(images, num_classes,
                       is_training=True,
                       config=None,
                       current_step=None):
  """Build NASNet model for the Cifar Dataset."""
  hparams = cifar_config() if config is None else copy.deepcopy(config)
  _update_hparams(hparams, is_training)

  if tf.test.is_gpu_available() and hparams.data_format == 'NHWC':
    tf.logging.info('A GPU is available on the machine, consider using NCHW '
                    'data format for increased speed on GPU.')

  if hparams.data_format == 'NCHW':
    images = tf.transpose(images, [0, 3, 1, 2])

  # Calculate the total number of cells in the network
  # Add 2 for the reduction cells
  total_num_cells = hparams.num_cells + 2

  normal_cell = nasnet_utils.NasNetANormalCell(
      hparams.num_conv_filters, hparams.drop_path_keep_prob,
      total_num_cells, hparams.total_training_steps)
  reduction_cell = nasnet_utils.NasNetAReductionCell(
      hparams.num_conv_filters, hparams.drop_path_keep_prob,
      total_num_cells, hparams.total_training_steps)
  with arg_scope([slim.dropout, nasnet_utils.drop_path, slim.batch_norm],
                 is_training=is_training):
    with arg_scope([slim.avg_pool2d,
                    slim.max_pool2d,
                    slim.conv2d,
                    slim.batch_norm,
                    slim.separable_conv2d,
                    nasnet_utils.factorized_reduction,
                    nasnet_utils.global_avg_pool,
                    nasnet_utils.get_channel_index,
                    nasnet_utils.get_channel_dim],
                   data_format=hparams.data_format):
      return _build_nasnet_base(images,
                                normal_cell=normal_cell,
                                reduction_cell=reduction_cell,
                                num_classes=num_classes,
                                hparams=hparams,
                                is_training=is_training,
                                stem_type='cifar',
                                current_step=current_step)
build_nasnet_cifar.default_image_size = 32


def build_nasnet_mobile(images, num_classes,
                        is_training=True,
                        final_endpoint=None,
                        config=None,
                        current_step=None):
  """Build NASNet Mobile model for the ImageNet Dataset."""
  hparams = (mobile_imagenet_config() if config is None
             else copy.deepcopy(config))
  _update_hparams(hparams, is_training)

  if tf.test.is_gpu_available() and hparams.data_format == 'NHWC':
    tf.logging.info('A GPU is available on the machine, consider using NCHW '
                    'data format for increased speed on GPU.')

  if hparams.data_format == 'NCHW':
    images = tf.transpose(images, [0, 3, 1, 2])

  # Calculate the total number of cells in the network
  # Add 2 for the reduction cells
  total_num_cells = hparams.num_cells + 2
  # If ImageNet, then add an additional two for the stem cells
  total_num_cells += 2

  normal_cell = nasnet_utils.NasNetANormalCell(
      hparams.num_conv_filters, hparams.drop_path_keep_prob,
      total_num_cells, hparams.total_training_steps)
  reduction_cell = nasnet_utils.NasNetAReductionCell(
      hparams.num_conv_filters, hparams.drop_path_keep_prob,
      total_num_cells, hparams.total_training_steps)
  with arg_scope([slim.dropout, nasnet_utils.drop_path, slim.batch_norm],
                 is_training=is_training):
    with arg_scope([slim.avg_pool2d,
                    slim.max_pool2d,
                    slim.conv2d,
                    slim.batch_norm,
                    slim.separable_conv2d,
                    nasnet_utils.factorized_reduction,
                    nasnet_utils.global_avg_pool,
                    nasnet_utils.get_channel_index,
                    nasnet_utils.get_channel_dim],
                   data_format=hparams.data_format):
      return _build_nasnet_base(images,
                                normal_cell=normal_cell,
                                reduction_cell=reduction_cell,
                                num_classes=num_classes,
                                hparams=hparams,
                                is_training=is_training,
                                stem_type='imagenet',
                                final_endpoint=final_endpoint,
                                current_step=current_step)
build_nasnet_mobile.default_image_size = 224


def build_nasnet_large(images, num_classes,
                       is_training=True,
                       final_endpoint=None,
                       config=None,
                       current_step=None):
  """Build NASNet Large model for the ImageNet Dataset."""
  hparams = (large_imagenet_config() if config is None
             else copy.deepcopy(config))
  _update_hparams(hparams, is_training)

  if tf.test.is_gpu_available() and hparams.data_format == 'NHWC':
    tf.logging.info('A GPU is available on the machine, consider using NCHW '
                    'data format for increased speed on GPU.')

  if hparams.data_format == 'NCHW':
    images = tf.transpose(images, [0, 3, 1, 2])

  # Calculate the total number of cells in the network
  # Add 2 for the reduction cells
  total_num_cells = hparams.num_cells + 2
  # If ImageNet, then add an additional two for the stem cells
  total_num_cells += 2

  normal_cell = nasnet_utils.NasNetANormalCell(
      hparams.num_conv_filters, hparams.drop_path_keep_prob,
      total_num_cells, hparams.total_training_steps)
  reduction_cell = nasnet_utils.NasNetAReductionCell(
      hparams.num_conv_filters, hparams.drop_path_keep_prob,
      total_num_cells, hparams.total_training_steps)
  with arg_scope([slim.dropout, nasnet_utils.drop_path, slim.batch_norm],
                 is_training=is_training):
    with arg_scope([slim.avg_pool2d,
                    slim.max_pool2d,
                    slim.conv2d,
                    slim.batch_norm,
                    slim.separable_conv2d,
                    nasnet_utils.factorized_reduction,
                    nasnet_utils.global_avg_pool,
                    nasnet_utils.get_channel_index,
                    nasnet_utils.get_channel_dim],
                   data_format=hparams.data_format):
      return _build_nasnet_base(images,
                                normal_cell=normal_cell,
                                reduction_cell=reduction_cell,
                                num_classes=num_classes,
                                hparams=hparams,
                                is_training=is_training,
                                stem_type='imagenet',
                                final_endpoint=final_endpoint,
                                current_step=current_step)
build_nasnet_large.default_image_size = 331


def _build_nasnet_base(images,
                       normal_cell,
                       reduction_cell,
                       num_classes,
                       hparams,
                       is_training,
                       stem_type,
                       final_endpoint=None,
                       current_step=None):
  """Constructs a NASNet image model."""

  end_points = {}
  def add_and_check_endpoint(endpoint_name, net):
    end_points[endpoint_name] = net
    return final_endpoint and (endpoint_name == final_endpoint)

  # Find where to place the reduction cells or stride normal cells
  reduction_indices = nasnet_utils.calc_reduction_layers(
      hparams.num_cells, hparams.num_reduction_layers)
  stem_cell = reduction_cell

  if stem_type == 'imagenet':
    stem = lambda: _imagenet_stem(images, hparams, stem_cell)
  elif stem_type == 'cifar':
    stem = lambda: _cifar_stem(images, hparams)
  else:
    raise ValueError('Unknown stem_type: ', stem_type)
  net, cell_outputs = stem()
  if add_and_check_endpoint('Stem', net): return net, end_points

  # Setup for building in the auxiliary head.
  aux_head_cell_idxes = []
  if len(reduction_indices) >= 2:
    aux_head_cell_idxes.append(reduction_indices[1] - 1)

  # Run the cells
  filter_scaling = 1.0
  # true_cell_num accounts for the stem cells
  true_cell_num = 2 if stem_type == 'imagenet' else 0
  for cell_num in range(hparams.num_cells):
    stride = 1
    if hparams.skip_reduction_layer_input:
      prev_layer = cell_outputs[-2]
    if cell_num in reduction_indices:
      filter_scaling *= hparams.filter_scaling_rate
      net = reduction_cell(
          net,
          scope='reduction_cell_{}'.format(reduction_indices.index(cell_num)),
          filter_scaling=filter_scaling,
          stride=2,
          prev_layer=cell_outputs[-2],
          cell_num=true_cell_num,
          current_step=current_step)
      if add_and_check_endpoint(
          'Reduction_Cell_{}'.format(reduction_indices.index(cell_num)), net):
        return net, end_points
      true_cell_num += 1
      cell_outputs.append(net)
    if not hparams.skip_reduction_layer_input:
      prev_layer = cell_outputs[-2]
    net = normal_cell(
        net,
        scope='cell_{}'.format(cell_num),
        filter_scaling=filter_scaling,
        stride=stride,
        prev_layer=prev_layer,
        cell_num=true_cell_num,
        current_step=current_step)

    if add_and_check_endpoint('Cell_{}'.format(cell_num), net):
      return net, end_points
    true_cell_num += 1
    if (hparams.use_aux_head and cell_num in aux_head_cell_idxes and
        num_classes and is_training):
      aux_net = tf.nn.relu(net)
      _build_aux_head(aux_net, end_points, num_classes, hparams,
                      scope='aux_{}'.format(cell_num))
    cell_outputs.append(net)

  # Final softmax layer
  with tf.variable_scope('final_layer'):
    net = tf.nn.relu(net)
    net = nasnet_utils.global_avg_pool(net)
    if add_and_check_endpoint('global_pool', net) or not num_classes:
      return net, end_points
    net = slim.dropout(net, hparams.dense_dropout_keep_prob, scope='dropout')
    logits = slim.fully_connected(net, num_classes)

    if add_and_check_endpoint('Logits', logits):
      return net, end_points

    predictions = tf.nn.softmax(logits, name='predictions')
    if add_and_check_endpoint('Predictions', predictions):
      return net, end_points
  return logits, end_points