diffaug.py

# Copyright (c) 2020, Shengyu Zhao, Zhijian Liu, Ji Lin, Jun-Yan Zhu, and Song Han
# All rights reserved.
#
# Redistribution and use in source and binary forms, with or without
# modification, are permitted provided that the following conditions are met:
#
# * Redistributions of source code must retain the above copyright notice, this
#   list of conditions and the following disclaimer.
#
# * Redistributions in binary form must reproduce the above copyright notice,
#   this list of conditions and the following disclaimer in the documentation
#   and/or other materials provided with the distribution.
#
# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
# DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
# FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
# DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
# SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
# CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
# OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
"""Differentiable Augmentation for Tensorflow.
Reference:
  - [Differentiable Augmentation for Data-Efficient GAN Training](
      https://arxiv.org/abs/2006.10738) (NeurIPS 2020)
"""

import tensorflow as tf


def DiffAugment(x, policy='', channels_first=False):
    if policy:
        if channels_first:
            x = tf.transpose(x, [0, 2, 3, 1])
        for p in policy.split(','):
            for f in AUGMENT_FNS[p]:
                x = f(x)
        if channels_first:
            x = tf.transpose(x, [0, 3, 1, 2])
    return x


def rand_brightness(x):
    magnitude = tf.random.uniform([tf.shape(x)[0], 1, 1, 1]) - 0.5
    x = x + magnitude
    return x


def rand_saturation(x):
    magnitude = tf.random.uniform([tf.shape(x)[0], 1, 1, 1]) * 2
    x_mean = tf.reduce_mean(x, axis=3, keepdims=True)
    x = (x - x_mean) * magnitude + x_mean
    return x


def rand_contrast(x):
    magnitude = tf.random.uniform([tf.shape(x)[0], 1, 1, 1]) + 0.5
    x_mean = tf.reduce_mean(x, axis=[1, 2, 3], keepdims=True)
    x = (x - x_mean) * magnitude + x_mean
    return x


def rand_translation(x, ratio=0.125):
    batch_size = tf.shape(x)[0]
    image_size = tf.shape(x)[1:3]
    shift = tf.cast(tf.cast(image_size, tf.float32) * ratio + 0.5, tf.int32)
    translation_x = tf.random.uniform([batch_size, 1], -shift[0], shift[0] + 1, dtype=tf.int32)
    translation_y = tf.random.uniform([batch_size, 1], -shift[1], shift[1] + 1, dtype=tf.int32)
    grid_x = tf.clip_by_value(tf.expand_dims(tf.range(image_size[0], dtype=tf.int32), 0) + translation_x + 1, 0, image_size[0] + 1)
    grid_y = tf.clip_by_value(tf.expand_dims(tf.range(image_size[1], dtype=tf.int32), 0) + translation_y + 1, 0, image_size[1] + 1)
    x = tf.gather_nd(tf.pad(x, [[0, 0], [1, 1], [0, 0], [0, 0]]), tf.expand_dims(grid_x, -1), batch_dims=1)
    x = tf.transpose(tf.gather_nd(tf.pad(tf.transpose(x, [0, 2, 1, 3]), [[0, 0], [1, 1], [0, 0], [0, 0]]), tf.expand_dims(grid_y, -1), batch_dims=1), [0, 2, 1, 3])
    return x


def rand_cutout(x, ratio=0.5):
    if tf.random.uniform([], minval=0.0, maxval=1.0) < 0.3:
        batch_size = tf.shape(x)[0]
        image_size = tf.shape(x)[1:3]
        cutout_size = tf.cast(tf.cast(image_size, tf.float32) * ratio + 0.5, tf.int32)
        offset_x = tf.random.uniform([tf.shape(x)[0], 1, 1], maxval=image_size[0] + (1 - cutout_size[0] % 2), dtype=tf.int32)
        offset_y = tf.random.uniform([tf.shape(x)[0], 1, 1], maxval=image_size[1] + (1 - cutout_size[1] % 2), dtype=tf.int32)
        grid_batch, grid_x, grid_y = tf.meshgrid(tf.range(batch_size, dtype=tf.int32), tf.range(cutout_size[0], dtype=tf.int32), tf.range(cutout_size[1], dtype=tf.int32), indexing='ij')
        cutout_grid = tf.stack([grid_batch, grid_x + offset_x - cutout_size[0] // 2, grid_y + offset_y - cutout_size[1] // 2], axis=-1)
        mask_shape = tf.stack([batch_size, image_size[0], image_size[1]])
        cutout_grid = tf.maximum(cutout_grid, 0)
        cutout_grid = tf.minimum(cutout_grid, tf.reshape(mask_shape - 1, [1, 1, 1, 3]))
        mask = tf.maximum(1 - tf.scatter_nd(cutout_grid, tf.ones([batch_size, cutout_size[0], cutout_size[1]], dtype=tf.float32), mask_shape), 0)
        x = x * tf.expand_dims(mask, axis=3)
    return x


AUGMENT_FNS = {
    'color': [rand_brightness, rand_saturation, rand_contrast],
    'translation': [rand_translation],
    'cutout': [rand_cutout],
}