custom_ops.py

# Copyright 2018 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 convenience wrappers for typical Neural Network TensorFlow layers.

   Ops that have different behavior during training or eval have an is_training
   parameter.
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function


import numpy as np
import tensorflow as tf


arg_scope = tf.contrib.framework.arg_scope


def variable(name, shape, dtype, initializer, trainable):
  """Returns a TF variable with the passed in specifications."""
  var = tf.get_variable(
      name,
      shape=shape,
      dtype=dtype,
      initializer=initializer,
      trainable=trainable)
  return var


def global_avg_pool(x, scope=None):
  """Average pools away spatial height and width dimension of 4D tensor."""
  assert x.get_shape().ndims == 4
  with tf.name_scope(scope, 'global_avg_pool', [x]):
    kernel_size = (1, int(x.shape[1]), int(x.shape[2]), 1)
    squeeze_dims = (1, 2)
    result = tf.nn.avg_pool(
        x,
        ksize=kernel_size,
        strides=(1, 1, 1, 1),
        padding='VALID',
        data_format='NHWC')
    return tf.squeeze(result, squeeze_dims)


def zero_pad(inputs, in_filter, out_filter):
  """Zero pads `input` tensor to have `out_filter` number of filters."""
  outputs = tf.pad(inputs, [[0, 0], [0, 0], [0, 0],
                            [(out_filter - in_filter) // 2,
                             (out_filter - in_filter) // 2]])
  return outputs


@tf.contrib.framework.add_arg_scope
def batch_norm(inputs,
               decay=0.999,
               center=True,
               scale=False,
               epsilon=0.001,
               is_training=True,
               reuse=None,
               scope=None):
  """Small wrapper around tf.contrib.layers.batch_norm."""
  return tf.contrib.layers.batch_norm(
      inputs,
      decay=decay,
      center=center,
      scale=scale,
      epsilon=epsilon,
      activation_fn=None,
      param_initializers=None,
      updates_collections=tf.GraphKeys.UPDATE_OPS,
      is_training=is_training,
      reuse=reuse,
      trainable=True,
      fused=True,
      data_format='NHWC',
      zero_debias_moving_mean=False,
      scope=scope)


def stride_arr(stride_h, stride_w):
  return [1, stride_h, stride_w, 1]


@tf.contrib.framework.add_arg_scope
def conv2d(inputs,
           num_filters_out,
           kernel_size,
           stride=1,
           scope=None,
           reuse=None):
  """Adds a 2D convolution.

  conv2d creates a variable called 'weights', representing the convolutional
  kernel, that is convolved with the input.

  Args:
    inputs: a 4D tensor in NHWC format.
    num_filters_out: the number of output filters.
    kernel_size: an int specifying the kernel height and width size.
    stride: an int specifying the height and width stride.
    scope: Optional scope for variable_scope.
    reuse: whether or not the layer and its variables should be reused.
  Returns:
    a tensor that is the result of a convolution being applied to `inputs`.
  """
  with tf.variable_scope(scope, 'Conv', [inputs], reuse=reuse):
    num_filters_in = int(inputs.shape[3])
    weights_shape = [kernel_size, kernel_size, num_filters_in, num_filters_out]

    # Initialization
    n = int(weights_shape[0] * weights_shape[1] * weights_shape[3])
    weights_initializer = tf.random_normal_initializer(
        stddev=np.sqrt(2.0 / n))

    weights = variable(
        name='weights',
        shape=weights_shape,
        dtype=tf.float32,
        initializer=weights_initializer,
        trainable=True)
    strides = stride_arr(stride, stride)
    outputs = tf.nn.conv2d(
        inputs, weights, strides, padding='SAME', data_format='NHWC')
    return outputs


@tf.contrib.framework.add_arg_scope
def fc(inputs,
       num_units_out,
       scope=None,
       reuse=None):
  """Creates a fully connected layer applied to `inputs`.

  Args:
    inputs: a tensor that the fully connected layer will be applied to. It
      will be reshaped if it is not 2D.
    num_units_out: the number of output units in the layer.
    scope: Optional scope for variable_scope.
    reuse: whether or not the layer and its variables should be reused.

  Returns:
     a tensor that is the result of applying a linear matrix to `inputs`.
  """
  if len(inputs.shape) > 2:
    inputs = tf.reshape(inputs, [int(inputs.shape[0]), -1])

  with tf.variable_scope(scope, 'FC', [inputs], reuse=reuse):
    num_units_in = inputs.shape[1]
    weights_shape = [num_units_in, num_units_out]
    unif_init_range = 1.0 / (num_units_out)**(0.5)
    weights_initializer = tf.random_uniform_initializer(
        -unif_init_range, unif_init_range)
    weights = variable(
        name='weights',
        shape=weights_shape,
        dtype=tf.float32,
        initializer=weights_initializer,
        trainable=True)
    bias_initializer = tf.constant_initializer(0.0)
    biases = variable(
        name='biases',
        shape=[num_units_out,],
        dtype=tf.float32,
        initializer=bias_initializer,
        trainable=True)
    outputs = tf.nn.xw_plus_b(inputs, weights, biases)
    return outputs


@tf.contrib.framework.add_arg_scope
def avg_pool(inputs, kernel_size, stride=2, padding='VALID', scope=None):
  """Wrapper around tf.nn.avg_pool."""
  with tf.name_scope(scope, 'AvgPool', [inputs]):
    kernel = stride_arr(kernel_size, kernel_size)
    strides = stride_arr(stride, stride)
    return tf.nn.avg_pool(
        inputs,
        ksize=kernel,
        strides=strides,
        padding=padding,
        data_format='NHWC')