fgsmutils.py

# coding=utf-8

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

import os
import numpy as np
import cv2
import pandas as pd
import scipy.stats as st
from scipy.misc import imread, imsave
from tensorflow.contrib.image import transform as images_transform
from tensorflow.contrib.image import rotate as images_rotate

import tensorflow as tf

from nets import inception_v3, inception_v4, inception_resnet_v2, resnet_v2

import random

from config import *


slim = tf.contrib.slim

tf.flags.DEFINE_integer('batch_size', 10, 'How many images process at one time.')

tf.flags.DEFINE_float('max_epsilon', 10.0, 'max epsilon.')

tf.flags.DEFINE_integer('num_iter', 10, 'max iteration.')

tf.flags.DEFINE_float('momentum', 1.0, 'momentum about the model.')

tf.flags.DEFINE_integer(
    'image_width', 299, 'Width of each input images.')

tf.flags.DEFINE_integer(
    'image_height', 299, 'Height of each input images.')

tf.flags.DEFINE_float('prob', 0.5, 'probability of using diverse inputs.')

tf.flags.DEFINE_integer('image_resize', 331, 'Height of each input images.')

tf.flags.DEFINE_string('checkpoint_path', './models',
                       'Path to checkpoint for pretained models.')

tf.flags.DEFINE_string('input_dir', './dev_data/val_rs',
                       'Input directory with images.')

tf.flags.DEFINE_string('output_dir', './outputs',
                       'Output directory with images.')

FLAGS = tf.flags.FLAGS

def load_images(input_dir, batch_shape):
    """Read png images from input directory in batches.
    Args:
      input_dir: input directory
      batch_shape: shape of minibatch array, i.e. [batch_size, height, width, 3]
    Yields:
      filenames: list file names without path of each image
        Lenght of this list could be less than batch_size, in this case only
        first few images of the result are elements of the minibatch.
      images: array with all images from this batch
    """
    images = np.zeros(batch_shape)
    filenames = []
    idx = 0
    batch_size = batch_shape[0]
    for filepath in tf.gfile.Glob(os.path.join(input_dir, '*'))[STARTING_INDEX:END_INDEX]:
        with tf.gfile.Open(filepath, 'rb') as f:
            image = imread(f, mode='RGB').astype(np.float) / 255.0
        # Images for inception classifier are normalized to be in [-1, 1] interval.
        images[idx, :, :, :] = image * 2.0 - 1.0
        filenames.append(os.path.basename(filepath))
        idx += 1
        if idx == batch_size:
            yield filenames, images
            filenames = []
            images = np.zeros(batch_shape)
            idx = 0
    if idx > 0:
        yield filenames, images


def save_images(images, filenames, output_dir):
    """Saves images to the output directory.

    Args:
        images: array with minibatch of images
        filenames: list of filenames without path
            If number of file names in this list less than number of images in
            the minibatch then only first len(filenames) images will be saved.
        output_dir: directory where to save images
    """
    for i, filename in enumerate(filenames):
        # Images for inception classifier are normalized to be in [-1, 1] interval,
        # so rescale them back to [0, 1].
        with tf.gfile.Open(os.path.join(output_dir, filename), 'w') as f:
            imsave(f, (images[i, :, :, :] + 1.0) * 0.5, format='png')


def check_or_create_dir(directory):
    """Check if directory exists otherwise create it."""
    if not os.path.exists(directory):
        os.makedirs(directory)


def graph(x, y, i, x_max, x_min, grad):
    eps = 2.0 * FLAGS.max_epsilon / 255.0
    num_iter = FLAGS.num_iter
    alpha = eps / num_iter
    momentum = FLAGS.momentum
    num_classes = 1001

    with slim.arg_scope(inception_v3.inception_v3_arg_scope()):
        logits_v3, end_points_v3 = inception_v3.inception_v3(
            x, num_classes=num_classes, is_training=False, reuse=tf.AUTO_REUSE)

    pred = tf.argmax(end_points_v3['Predictions'], 1)

    first_round = tf.cast(tf.equal(i, 0), tf.int64)
    y = first_round * pred + (1 - first_round) * y
    one_hot = tf.one_hot(y, num_classes)

    cross_entropy = tf.losses.softmax_cross_entropy(one_hot, logits_v3)
    noise = tf.gradients(cross_entropy, x)[0]

    noise = noise / tf.reduce_mean(tf.abs(noise), [1, 2, 3], keep_dims=True)

    noise = momentum * grad + noise

    x = x + alpha * tf.sign(noise)
    x = tf.clip_by_value(x, x_min, x_max)
    i = tf.add(i, 1)

    return x, y, i, x_max, x_min, noise


def stop(x, y, i, x_max, x_min, grad):
    num_iter = FLAGS.num_iter
    return tf.less(i, num_iter)


def image_augmentation(x):
    # img, noise
    one = tf.fill([tf.shape(x)[0], 1], 1.)
    zero = tf.fill([tf.shape(x)[0], 1], 0.)
    transforms = tf.concat([one, zero, zero, zero, one, zero, zero, zero], axis=1)
    rands = tf.concat([tf.truncated_normal([tf.shape(x)[0], 6], stddev=0.05), zero, zero], axis=1)
    return images_transform(x, transforms + rands, interpolation='BILINEAR')


def image_rotation(x):
    """ imgs, scale, scale is in radians """
    rands = tf.truncated_normal([tf.shape(x)[0]], stddev=0.05)
    return images_rotate(x, rands, interpolation='BILINEAR')


def input_diversity(input_tensor):
    rnd = tf.random_uniform((), FLAGS.image_width, FLAGS.image_resize, dtype=tf.int32)
    rescaled = tf.image.resize_images(input_tensor, [rnd, rnd], method=tf.image.ResizeMethod.NEAREST_NEIGHBOR)
    h_rem = FLAGS.image_resize - rnd
    w_rem = FLAGS.image_resize - rnd
    pad_top = tf.random_uniform((), 0, h_rem, dtype=tf.int32)
    pad_bottom = h_rem - pad_top
    pad_left = tf.random_uniform((), 0, w_rem, dtype=tf.int32)
    pad_right = w_rem - pad_left
    padded = tf.pad(rescaled, [[0, 0], [pad_top, pad_bottom], [pad_left, pad_right], [0, 0]], constant_values=0.)
    padded.set_shape((input_tensor.shape[0], FLAGS.image_resize, FLAGS.image_resize, 3))
    ret = tf.cond(tf.random_uniform(shape=[1])[0] < tf.constant(FLAGS.prob), lambda: padded, lambda: input_tensor)
    ret = tf.image.resize_images(ret, [FLAGS.image_height, FLAGS.image_width],
                                 method=tf.image.ResizeMethod.NEAREST_NEIGHBOR)
    return ret
    # return tf.cond(tf.random_uniform(shape=[1])[0] < tf.constant(FLAGS.prob), lambda: padded, lambda: input_tensor)

def load_labels(file_name):
    import pandas as pd
    dev = pd.read_csv(file_name)
    f2l = {dev.iloc[i]['filename']: dev.iloc[i]['label'] for i in range(len(dev))}
    return f2l