pix2pix_w_hint_lab_wgan_larger_sketch_mix_cond_refactored_dilation_deeper.py

#!/usr/bin/env python
# -*- coding: utf-8 -*-
"""
This file implements the main framework for the Sketch Coloring program. It takes colored images side by side with
their corresponding sketches and trains to convert one to the other. It is a tensorflow implementation of the
PaintsChainer project (https://paintschainer.preferred.tech/) with modifications in some details.
This implementation's framework mainly came from the pix2pix project https://github.com/affinelayer/pix2pix-tensorflow .

Example usage
TODO: change those example usages.
# Sanity check
python pix2pix_w_hint_lab_wgan_larger_sketch_mix.py --mode train --output_dir sanity_check --max_epochs 2000 --input_dir sanity_check_images --which_direction AtoB --display_freq=200 --gray_input_a --batch_size 1 --lr 0.0008 --gpu_percentage 0.1 --scale_size=143 --crop_size=128 --use_sketch_loss --pretrained_sketch_net_path sketch_colored_pair_cleaned_128_w_hint_lab_wgan_larger_train_sketch --use_hint --lab_colorization
# Train
python pix2pix_w_hint_lab_wgan_larger_sketch_mix.py --mode train --output_dir pixiv_downloaded_128_w_hint_lab_wgan_larger_sketch_mix --max_epochs 20 --input_dir /mnt/data_drive/home/ubuntu/pixiv_downloaded_sketches_lnet_128_combined/train --which_direction AtoB --display_freq=1000 --gray_input_a --batch_size 4 --lr 0.0008 --gpu_percentage 0.75 --scale_size=143 --crop_size=128 --use_sketch_loss --pretrained_sketch_net_path sketch_colored_pair_cleaned_128_w_hint_lab_wgan_larger_train_sketch --use_hint --lab_colorization
# Train 512
python pix2pix_w_hint_lab_wgan_larger_sketch_mix.py --mode train --output_dir pixiv_downloaded_512_w_hint_lab_wgan_larger_sketch_mix --max_epochs 20 --input_dir /mnt/data_drive/home/ubuntu/pixiv_downloaded_sketches_lnet_512_combined/train --which_direction AtoB --display_freq=1000 --gray_input_a --batch_size 4 --lr 0.0008 --gpu_percentage 0.9 --scale_size=572 --crop_size=512 --use_sketch_loss --pretrained_sketch_net_path pixiv_full_128_to_sketch_train --use_hint --lab_colorization --checkpoint=pixiv_downloaded_512_w_hint_lab_wgan_larger_sketch_mix --from_128
# Test
python pix2pix_w_hint_lab_wgan_larger_sketch_mix.py --mode test --output_dir pixiv_downloaded_128_w_hint_lab_wgan_larger_sketch_mix_test_with_hint --max_epochs 20 --input_dir /mnt/data_drive/home/ubuntu/pixiv_downloaded_sketches_lnet_128_combined/test --which_direction AtoB --display_freq=1000 --gray_input_a --batch_size 4 --lr 0.0008 --gpu_percentage 0.45 --scale_size=143 --crop_size=128 --use_sketch_loss --pretrained_sketch_net_path pixiv_full_128_to_sketch_train --use_hint --lab_colorization --checkpoint=pixiv_downloaded_128_w_hint_lab_wgan_larger_sketch_mix
python pix2pix_w_hint_lab_wgan_larger_sketch_mix.py --mode test --output_dir pixiv_downloaded_128_w_hint_lab_wgan_larger_sketch_mix_test_no_hint --max_epochs 20 --input_dir /mnt/data_drive/home/ubuntu/pixiv_downloaded_sketches_lnet_128_combined/test --which_direction AtoB --display_freq=1000 --gray_input_a --batch_size 4 --lr 0.0008 --gpu_percentage 0.45 --scale_size=143 --crop_size=128 --use_sketch_loss --pretrained_sketch_net_path pixiv_full_128_to_sketch_train --use_hint --lab_colorization --checkpoint=pixiv_downloaded_128_w_hint_lab_wgan_larger_sketch_mix --user_hint_path=BLANK

# To train a network that turns colored images into sketches:
python pix2pix_w_hint_lab_wgan_larger_sketch_mix.py --mode train --output_dir sketch_colored_pair_cleaned_128_w_hint_lab_wgan_larger_sketch_mix_train_sketch --max_epochs 20 --input_dir /mnt/data_drive/home/ubuntu/sketch_colored_pair_128_combined_cleaned/sketch_colored_pair_128_combined/train/ --which_direction BtoA --display_freq=1000 --gray_input_a --batch_size 4 --lr 0.0008 --gpu_percentage 0.45 --scale_size=143 --crop_size=128 --lab_colorization --train_sketch
# Create the a new sketch database using the pretrained sketch generation network.
python pix2pix_w_hint_lab_wgan_larger_sketch_mix.py --mode test --output_dir /mnt/data_drive/home/ubuntu/pixiv_downloaded_sketches_lnet_128/new_sketch --max_epochs 20 --input_dir /mnt/data_drive/home/ubuntu/pixiv_downloaded_sketches_lnet_128/color/ --which_direction BtoA --display_freq=1000 --gray_input_a --batch_size 4 --lr 0.0008 --gpu_percentage 0.25 --scale_size=143 --crop_size=128 --lab_colorization --train_sketch --checkpoint=sketch_colored_pair_cleaned_128_w_hint_lab_wgan_larger_sketch_mix_train_sketch --single_input

"""

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

import tensorflow as tf
import numpy as np
import argparse
import os
import json
import glob
import random
import collections
import math
import time
import urllib

from general_util import imread, get_all_image_paths
from neural_util import decode_image, decode_image_with_file_name
from sketches_util import sketch_extractor

parser = argparse.ArgumentParser()
parser.add_argument("--input_dir", required=True, help="Path to folder containing images to be trained, "
                                                       "or a file containing paths to the images. "
                                                       "The images should come from running the preprocess program, or "
                                                       "they have to be sketch images on the left and the "
                                                       "corresponding colored images on the right.")
parser.add_argument("--mode", required=True, choices=["train", "test"])
parser.add_argument("--output_dir", required=True, help="where to put output files.")
parser.add_argument("--seed", type=int)
parser.add_argument("--checkpoint", default=None,
                    help="Directory with checkpoint to resume training from or use for testing")
parser.add_argument("--user_hint_path", default=None, help="Path to a hint image for testing.")
parser.add_argument("--pretrained_sketch_net_path", default=None,
                    help="Path to the pretrained sketch network checkpoint to be used by the sketch loss.")
parser.add_argument("--single_input", dest="single_input", action="store_true",
                    help="If set, the input image is a single image instead of a combination of the source and target.")
parser.set_defaults(single_input=False)
parser.add_argument("--output_ab", dest="output_ab", action="store_true",
                    help="The generator network outputs only ab channel instead of lab. "
                         "Must be used with lab_colorization. This is for testing grayscale image to colored image "
                         "conversion.")
parser.set_defaults(output_ab=False)
parser.add_argument("--gen_sketch_input", dest="gen_sketch_input", action="store_true",
                    help="If set, instead of taking the given sketch image as input, it will generate the input image "
                         "using the sketch generator network. Intended to be used with sketch loss on. "
                         "Either the mix probability should be set to 0 or the single_input should be on.")
parser.set_defaults(gen_sketch_input=False)
parser.add_argument("--max_steps", type=int, help="number of training steps (0 to disable)")
parser.add_argument("--max_epochs", type=int, help="number of training epochs")
parser.add_argument("--summary_freq", type=int, default=10, help="update summaries every summary_freq steps")
parser.add_argument("--progress_freq", type=int, default=50, help="display progress every progress_freq steps")
# to get tracing working on GPU, LD_LIBRARY_PATH may need to be modified:
# LD_LIBRARY_PATH=/usr/local/nvidia/lib:/usr/local/nvidia/lib64:/usr/local/cuda/extras/CUPTI/lib64
parser.add_argument("--trace_freq", type=int, default=0, help="trace execution every trace_freq steps")
parser.add_argument("--display_freq", type=int, default=0, help="write current training images every display_freq steps")
parser.add_argument("--save_freq", type=int, default=2500, help="save model every save_freq steps, 0 to disable")

parser.add_argument("--aspect_ratio", type=float, default=1.0, help="aspect ratio of output images (width/height)")
parser.add_argument("--lab_colorization", action="store_true",
                    help="If set, it trains images in lab color space instead of rgb.")
parser.add_argument("--gray_input_a", action="store_true", help="Treat A image as grayscale image.")
parser.add_argument("--gray_input_b", action="store_true", help="Treat B image as grayscale image.")
parser.add_argument("--use_hint", action="store_true",
                    help="Supply random hint pixels taken from the target as extra input channels.")
parser.add_argument("--batch_size", type=int, default=1, help="number of images in batch")
parser.add_argument("--which_direction", type=str, default="BtoA", choices=["AtoB", "BtoA"],
                    help="BtoA is for colored to sketch images. AtoB is for the other way around.")
parser.add_argument("--ngf", type=int, default=32, help="number of generator filters in the first conv layer")
parser.add_argument("--ndf", type=int, default=32, help="number of discriminator filters in the first conv layer")
parser.add_argument("--scale_size", type=int, default= 572,
                    help="scale images to this size before cropping to `CROP_SIZE`x`CROP_SIZE`")
parser.add_argument("--crop_size", type=int, default= 512,
                    help="Crop scaled input images at random positions to this size.")
parser.add_argument("--flip", dest="flip", action="store_true", help="Randomly flip images horizontally during training.")
parser.add_argument("--no_flip", dest="flip", action="store_false", help="Disable random flipping during training.")
parser.set_defaults(flip=True)
parser.add_argument("--lr", type=float, default=0.0002, help="initial learning rate")
# beta1 not used because we're using RMS to be compatible with WGAN.
# parser.add_argument("--beta1", type=float, default=0.5, help="momentum term of adam")
parser.add_argument("--l1_weight", type=float, default=100.0, help="weight on L1 term for generator gradient")
parser.add_argument("--gan_weight", type=float, default=1.0, help="weight on GAN term for generator gradient")
parser.add_argument("--sketch_weight", type=float, default=1.0, help="weight on sketch loss term.")
parser.add_argument("--gpu_percentage", type=float, default=0.45,
                    help="The percentage of gpu this program can use. Set to <= 0 for cpu mode.")
parser.add_argument("--from_128", dest="from_128", action="store_true",
                    help="Indicate whether model is from a previous model trained on 128x128 images. It should be used "
                         "when the pretrained model needs to be retrained on a different input size. The model "
                         "can inherit and reuse most of the parameters from the provided pretrained model.")
parser.add_argument("--train_sketch", dest="train_sketch", action="store_true",
                    help="Indicate whether the model is for sketch generation. Variable scope will change accordingly.")
parser.add_argument("--use_sketch_loss", dest="use_sketch_loss", action="store_true",
                    help="Use the pretrained sketch generator network to compare the sketches of the generated image "
                         "versus that of the original image. the `pretrained_sketch_net_path` must be provided.")
parser.add_argument("--hint_prob", type=float, default=0.5,
                    help="The probability of providing hints as extra input channels.")
parser.add_argument("--mix_prob", type=float, default=0.5,
                    help="The probability of having sketch generated by dilation as the input instead of using the "
                         "provided input sketches.")

a = parser.parse_args()
# assert a.mix_prob >= 1

# if a.use_sketch_loss and a.pretrained_sketch_net_path is None:
#     parser.error("If you want to use sketch loss, please provide a valid pretrained_sketch_net_path.")
if a.gen_sketch_input and not a.use_sketch_loss:
    parser.error("If you want to use gen_sketch_input, please also turn on sketch loss.")
if a.gen_sketch_input and a.mix_prob > 0:
    parser.error("If you want to use gen_sketch_input, please set mix_prob to 0 or below.")
if a.mode != "test" and a.single_input and (a.mix_prob < 1 and not a.gen_sketch_input):
    parser.error("single input mode is intended for either test mode where only output is needed, or training where the "
                 "sketch is generated simply by dilation.")
if a.output_ab and not a.lab_colorization :
    parser.error("If you want the generator to output only a and b channels, please also add lab_colorization flag.")
if a.sketch_weight != 10.0:
    raw_input("Are you sure you don't want sketch_weight to be 10.0?")

CROP_SIZE = a.crop_size  # 128 # 256
# The discriminator weights will be capped at the CLIP_VALUE.
# In the original WGAN paper it was 0.01, but that value was too small to provide colorful outputs.
CLIP_VALUE = 0.04
# The number of parameters to be trained may be different depending on the input size and the ngf and ndf values.
APPROXIMATE_NUMBER_OF_TOTAL_PARAMETERS = 19523936
SKETCH_VAR_SCOPE_PREFIX = "sketch_"

Examples = collections.namedtuple("Examples", "paths, inputs, targets, count, steps_per_epoch, input_hints, is_hint_off")
Model = collections.namedtuple("Model", "outputs, predict_real, predict_fake, real_sketches, fake_sketches, discrim_loss, gen_loss_GAN, gen_loss_L1, gen_loss_sketch, train")


def conv(batch_input, out_channels, stride, shift=4, pad = 1, trainable=True):
    with tf.variable_scope("conv"):
        in_channels = batch_input.get_shape()[3]
        filter = tf.get_variable("filter", [shift, shift, in_channels, out_channels], dtype=tf.float32, initializer=tf.random_normal_initializer(0, 0.02), trainable=trainable)
        # [batch, in_height, in_width, in_channels], [filter_width, filter_height, in_channels, out_channels]
        #     => [batch, out_height, out_width, out_channels]
        if pad > 0:
            padded_input = tf.pad(batch_input, [[0, 0], [pad, pad], [pad, pad], [0, 0]], mode="CONSTANT")
        else:
            assert pad == 0
            padded_input = batch_input
        conv = tf.nn.conv2d(padded_input, filter, [1, stride, stride, 1], padding="VALID")
        return conv


def lrelu(x, a):
    with tf.name_scope("lrelu"):
        # adding these together creates the leak part and linear part
        # then cancels them out by subtracting/adding an absolute value term
        # leak: a*x/2 - a*abs(x)/2
        # linear: x/2 + abs(x)/2

        # this block looks like it has 2 inputs on the graph unless we do this
        x = tf.identity(x)
        return (0.5 * (1 + a)) * x + (0.5 * (1 - a)) * tf.abs(x)


def batchnorm(input, trainable=True):
    with tf.variable_scope("batchnorm"):
        # this block looks like it has 3 inputs on the graph unless we do this
        input = tf.identity(input)

        channels = input.get_shape()[3]
        offset = tf.get_variable("offset", [channels], dtype=tf.float32, initializer=tf.zeros_initializer, trainable=trainable)
        scale = tf.get_variable("scale", [channels], dtype=tf.float32, initializer=tf.random_normal_initializer(1.0, 0.02), trainable=trainable)
        mean, variance = tf.nn.moments(input, axes=[0, 1, 2], keep_dims=False)
        variance_epsilon = 1e-5
        normalized = tf.nn.batch_normalization(input, mean, variance, offset, scale, variance_epsilon=variance_epsilon)
        return normalized


def deconv(batch_input, out_channels, stride = 2, shift = 4, trainable=True):
    with tf.variable_scope("deconv"):
        batch, in_height, in_width, in_channels = [int(d) for d in batch_input.get_shape()]
        filter = tf.get_variable("filter", [shift, shift, out_channels, in_channels], dtype=tf.float32, initializer=tf.random_normal_initializer(0, 0.02), trainable=trainable)
        # [batch, in_height, in_width, in_channels], [filter_width, filter_height, out_channels, in_channels]
        #     => [batch, out_height, out_width, out_channels]
        conv = tf.nn.conv2d_transpose(batch_input, filter, [batch, in_height * stride, in_width * stride, out_channels], [1, stride, stride, 1], padding="SAME")
        return conv


def check_image(image):
    assertion = tf.assert_equal(tf.shape(image)[-1], 3, message="image must have 3 color channels")
    with tf.control_dependencies([assertion]):
        image = tf.identity(image)

    if image.get_shape().ndims not in (3, 4):
        raise ValueError("image must be either 3 or 4 dimensions")

    # make the last dimension 3 so that you can unstack the colors
    shape = list(image.get_shape())
    shape[-1] = 3
    image.set_shape(shape)
    return image

# based on https://github.com/torch/image/blob/9f65c30167b2048ecbe8b7befdc6b2d6d12baee9/generic/image.c
def rgb_to_lab(srgb):
    # Input range [0, 255]
    with tf.name_scope("rgb_to_lab"):
        srgb = check_image(srgb)
        srgb_pixels = tf.reshape(srgb, [-1, 3])

        with tf.name_scope("srgb_to_xyz"):
            linear_mask = tf.cast(srgb_pixels <= 0.04045, dtype=tf.float32)
            exponential_mask = tf.cast(srgb_pixels > 0.04045, dtype=tf.float32)
            rgb_pixels = (srgb_pixels / 12.92 * linear_mask) + (((srgb_pixels + 0.055) / 1.055) ** 2.4) * exponential_mask
            rgb_to_xyz = tf.constant([
                #    X        Y          Z
                [0.412453, 0.212671, 0.019334], # R
                [0.357580, 0.715160, 0.119193], # G
                [0.180423, 0.072169, 0.950227], # B
            ])
            xyz_pixels = tf.matmul(rgb_pixels, rgb_to_xyz)

        # https://en.wikipedia.org/wiki/Lab_color_space#CIELAB-CIEXYZ_conversions
        with tf.name_scope("xyz_to_cielab"):
            # convert to fx = f(X/Xn), fy = f(Y/Yn), fz = f(Z/Zn)

            # normalize for D65 white point
            xyz_normalized_pixels = tf.multiply(xyz_pixels, [1/0.950456, 1.0, 1/1.088754])

            epsilon = 6/29
            linear_mask = tf.cast(xyz_normalized_pixels <= (epsilon**3), dtype=tf.float32)
            exponential_mask = tf.cast(xyz_normalized_pixels > (epsilon**3), dtype=tf.float32)
            fxfyfz_pixels = (xyz_normalized_pixels / (3 * epsilon**2) + 4/29) * linear_mask + (xyz_normalized_pixels ** (1/3)) * exponential_mask

            # convert to lab
            fxfyfz_to_lab = tf.constant([
                #  l       a       b
                [  0.0,  500.0,    0.0], # fx
                [116.0, -500.0,  200.0], # fy
                [  0.0,    0.0, -200.0], # fz
            ])
            lab_pixels = tf.matmul(fxfyfz_pixels, fxfyfz_to_lab) + tf.constant([-16.0, 0.0, 0.0])

        return tf.reshape(lab_pixels, tf.shape(srgb))

def lab_to_rgb(lab):
    # Input range for l is 0 ~ 100 and ab is -110 ~ 110
    # Output range is 0 ~ 1....???
    with tf.name_scope("lab_to_rgb"):
        lab = check_image(lab)
        lab_pixels = tf.reshape(lab, [-1, 3])

        # https://en.wikipedia.org/wiki/Lab_color_space#CIELAB-CIEXYZ_conversions
        with tf.name_scope("cielab_to_xyz"):
            # convert to fxfyfz
            lab_to_fxfyfz = tf.constant([
                #   fx      fy        fz
                [1/116.0, 1/116.0,  1/116.0], # l
                [1/500.0,     0.0,      0.0], # a
                [    0.0,     0.0, -1/200.0], # b
            ])
            fxfyfz_pixels = tf.matmul(lab_pixels + tf.constant([16.0, 0.0, 0.0]), lab_to_fxfyfz)

            # convert to xyz
            epsilon = 6/29
            linear_mask = tf.cast(fxfyfz_pixels <= epsilon, dtype=tf.float32)
            exponential_mask = tf.cast(fxfyfz_pixels > epsilon, dtype=tf.float32)
            xyz_pixels = (3 * epsilon**2 * (fxfyfz_pixels - 4/29)) * linear_mask + (fxfyfz_pixels ** 3) * exponential_mask

            # denormalize for D65 white point
            xyz_pixels = tf.multiply(xyz_pixels, [0.950456, 1.0, 1.088754])

        with tf.name_scope("xyz_to_srgb"):
            xyz_to_rgb = tf.constant([
                #     r           g          b
                [ 3.2404542, -0.9692660,  0.0556434], # x
                [-1.5371385,  1.8760108, -0.2040259], # y
                [-0.4985314,  0.0415560,  1.0572252], # z
            ])
            rgb_pixels = tf.matmul(xyz_pixels, xyz_to_rgb)
            # avoid a slightly negative number messing up the conversion
            rgb_pixels = tf.clip_by_value(rgb_pixels, 0.0, 1.0)
            linear_mask = tf.cast(rgb_pixels <= 0.0031308, dtype=tf.float32)
            exponential_mask = tf.cast(rgb_pixels > 0.0031308, dtype=tf.float32)
            srgb_pixels = (rgb_pixels * 12.92 * linear_mask) + ((rgb_pixels ** (1/2.4) * 1.055) - 0.055) * exponential_mask

        return tf.reshape(srgb_pixels, tf.shape(lab))

def load_examples(user_hint = None):
    if not os.path.exists(a.input_dir):
        raise Exception("input_dir does not exist")

    input_paths = get_all_image_paths(a.input_dir)
    decode = decode_image_with_file_name

    if len(input_paths) == 0:
        raise Exception("input_dir contains no image files")

    def get_name(path):
        name, _ = os.path.splitext(os.path.basename(path))
        return name

    # if the image names are numbers, sort by the value rather than asciibetically
    # having sorted inputs means that the outputs are sorted in test mode
    if all(get_name(path).isdigit() for path in input_paths):
        input_paths = sorted(input_paths, key=lambda path: int(get_name(path)))
    else:
        input_paths = sorted(input_paths)

    with tf.name_scope("load_images"):
        path_queue = tf.train.string_input_producer(input_paths, shuffle=a.mode == "train")
        reader = tf.WholeFileReader()
        paths, contents = reader.read(path_queue)
        raw_input = decode(contents, paths, channels=3)
        raw_input = tf.image.convert_image_dtype(raw_input, dtype=tf.float32)

        assertion = tf.assert_equal(tf.shape(raw_input)[2], 3, message="image does not have 3 channels")
        with tf.control_dependencies([assertion]):
            raw_input = tf.identity(raw_input)

        raw_input.set_shape([None, None, 3])

        # break apart image pair and move to range [-1, 1]
        width = tf.shape(raw_input)[1]  # [height, width, channels]

        # Note that a_images and b_images have ranges 0~1 before being fed into grayscale and rescaling.
        if a.single_input:
            a_images = raw_input
            b_images = raw_input
        else:
            a_images = raw_input[:, :width // 2, :]
            b_images = raw_input[:, width // 2:, :]
        if a.gray_input_a:
            a_images = tf.image.rgb_to_grayscale(a_images)
        if a.gray_input_b:
            b_images = tf.image.rgb_to_grayscale(b_images)

        # Note there is a subtle difference between first generating sketches on the larger image then rescaling them
        # (which is what this piece of code is for) and first rescale then generate sketches. First generating sketches
        # will result in inputs with much denser and messier looking sketches, which is not good when the scale is
        # small (like 128x128). This only matters when you have input images sizes different from `crop_size` and when
        # you're trying to train using the sketches generated from the target image using dilation.
        if a.mix_prob >= 1:
            a_images = sketch_extractor(b_images, color_space="rgb", max_val=1.0, min_val=0.0)
        elif a.mix_prob <= 0:
            a_images = a_images
        else:
            random_mix_condition = tf.random_uniform(shape=[], minval=0, maxval=1, dtype=tf.float32,
                                                     name="random_mix_condition")
            mix_prob = tf.constant(a.mix_prob)
            a_images = tf.cond(tf.greater_equal(random_mix_condition, mix_prob), lambda: a_images,
                               lambda: sketch_extractor(b_images, color_space="rgb", max_val=1.0, min_val=0.0))

        if a.lab_colorization:
            # Assume b image (the image on the right) is the colored image.
            lab = rgb_to_lab(b_images)
            L_chan, a_chan, b_chan = tf.unstack(lab, axis=2)

            L_chan = tf.expand_dims(L_chan, axis=2) / 50 - 1 # black and white with input range [0, 100]
            ab_chan = tf.stack([a_chan, b_chan], axis=2) / 110 # color channels with input range ~[-110, 110], not exact
            b_images = tf.concat(2,[L_chan, ab_chan])
            a_images = a_images * 2 - 1
        else:
            a_images = a_images * 2 - 1
            b_images = b_images * 2 - 1


    if a.which_direction == "AtoB":
        inputs, targets = [a_images, b_images]
    elif a.which_direction == "BtoA":
        inputs, targets = [b_images, a_images]
    else:
        raise Exception("invalid direction")

    # synchronize seed for image operations so that we do the same operations to both
    # input and output images
    seed = random.randint(0, 2**31 - 1)
    def transform(image):
        r = image
        if a.flip:
            r = tf.image.random_flip_left_right(r, seed=seed)
        # area produces a nice downscaling, but does nearest neighbor for upscaling
        # assume we're going to be doing downscaling here
        r = tf.image.resize_images(r, [a.scale_size, a.scale_size], method=tf.image.ResizeMethod.AREA)

        offset = tf.cast(tf.floor(tf.random_uniform([2], 0, a.scale_size - CROP_SIZE + 1, seed=seed)), dtype=tf.int32)
        if a.scale_size > CROP_SIZE:
            r = tf.image.crop_to_bounding_box(r, offset[0], offset[1], CROP_SIZE, CROP_SIZE)
        elif a.scale_size < CROP_SIZE:
            raise Exception("scale size cannot be less than crop size")
        return r

    def append_hint(inputs, targets, user_hint = None):
        num_hints = 40
        # Because the inputs are scaled -1~1, the blank hint should default to all -1s.
        blank_hint = np.ones(targets.get_shape().as_list()[:-1] + [4], dtype=np.float32) * -1
        output = tf.get_variable('output', initializer=blank_hint, dtype=tf.float32, trainable=False)

        if user_hint is None:
            rd_indices_h = tf.random_uniform([num_hints, 1], minval=0, maxval=targets.get_shape().as_list()[-3], dtype=tf.int32)
            rd_indices_w = tf.random_uniform([num_hints, 1], minval=0, maxval=targets.get_shape().as_list()[-2], dtype=tf.int32)
            rd_indices_2d = tf.concat(1,(rd_indices_h,rd_indices_w))

            targets_rgba = tf.concat(2,(targets,np.ones(targets.get_shape().as_list()[:-1] + [1])))
            hints = tf.gather_nd(targets_rgba, rd_indices_2d)
            clear_hint_op = tf.assign(output, blank_hint)
            random_condition = tf.random_uniform(shape=[], minval=0, maxval=1, dtype=tf.float32, name="random_hint_condition")
            with tf.control_dependencies([clear_hint_op,hints]):
                # I cannot assign this to any other variable, otherwise it will cause the program to be confused on
                # whether the clear hint op should be ran first or the scatter update first.
                half_constant = tf.constant(a.hint_prob)
                is_hint_off = tf.greater_equal(random_condition, half_constant, name="is_hint_off")
                output = tf.cond(is_hint_off, lambda: output, lambda: tf.scatter_nd_update(output, rd_indices_2d, hints))
        else:
            output = tf.assign(output, user_hint)

        assert len(output.get_shape().as_list()) == 3
        return tf.concat(2, (inputs, output), name='input_concat'), output, is_hint_off
        # return tf.concat(2, (inputs, output), name='input_concat'), output

    with tf.name_scope("target_images"):
        target_images = transform(targets)
    with tf.name_scope("input_images"):
        input_images = transform(inputs)

        # Please check the note above about the difference between rescale first and generate sketch first.
        # This code is kept here in case you need to switch to rescaling before generate sketches.
        # random_mix_condition = tf.random_uniform(shape=[], minval=0, maxval=1, dtype=tf.float32,
        #                                          name="random_mix_condition")
        # mix_prob = tf.constant(a.mix_prob)
        # input_images = tf.cond(tf.greater_equal(random_mix_condition, mix_prob), lambda: input_images,
        #                    lambda: sketch_extractor(target_images, max_val=1.0, min_val=-1.0, color_space="lab" if a.lab_colorization else "rgb"))

        if a.use_hint:
            input_images, input_hints, is_hint_off = append_hint(input_images, target_images, user_hint=user_hint)


    if a.use_hint:
        paths, inputs, targets, input_hints, is_hint_off = tf.train.batch([paths, input_images, target_images, input_hints, is_hint_off], batch_size=a.batch_size)
    else:
        paths, inputs, targets = tf.train.batch([paths, input_images, target_images], batch_size=a.batch_size)
    steps_per_epoch = int(math.ceil(len(input_paths) / a.batch_size))

    return Examples(
        paths=paths,
        inputs=inputs,
        targets=targets,
        count=len(input_paths),
        steps_per_epoch=steps_per_epoch,
        input_hints=input_hints if a.use_hint else None,
        is_hint_off=is_hint_off if a.use_hint else None,
    )


def create_model(inputs, targets, is_hint_off=None):
    def create_generator(generator_inputs, generator_outputs_channels, trainable = True):
        layers = []

        # encoder_1: [batch, h, w, in_channels] => [batch, h, w, ngf]
        with tf.variable_scope("encoder_1"):
            convolved = conv(generator_inputs, a.ngf, stride=1, shift=3, trainable=trainable)
            output = batchnorm(convolved, trainable=trainable)
            rectified = tf.nn.relu(output)  #TODO: maybe test leaky relu instead (like lrelu(output, 0.2))?
            layers.append(rectified)

        layer_specs = [
            a.ngf * 2, # encoder_2: [batch, h, w, ngf] => [batch, h/2, w/2, ngf * 2]
            a.ngf * 2, # encoder_3: [batch, h/2, w/2, ngf * 2] => [batch, h/2, w/2, ngf * 2]
            a.ngf * 4, # encoder_4: [batch, h/2, w/2, ngf * 2] => [batch, h/4, w/4, ngf * 4]
            a.ngf * 4, # encoder_5: [batch, h/4, w/4, ngf * 4] => [batch, h/4, w/4, ngf * 4]
            a.ngf * 8, # encoder_6: [batch, h/4, w/4, ngf * 4] => [batch, h/8, w/8, ngf * 8]
            a.ngf * 8, # encoder_7: [batch, h/8, w/8, ngf * 8] => [batch, h/8, w/8, ngf * 8]
            a.ngf * 16, # encoder_8: [batch, h/8, w/8, ngf * 8] => [batch, h/16, w/16, ngf * 16]
            a.ngf * 16, # encoder_9: [batch, h/16, w/16, ngf * 16] => [batch, h/16, w/16, ngf * 16]
            a.ngf * 16,  # encoder_10: [batch, h/16, w/16, ngf * 16] => [batch, h/32, w/32, ngf * 16]
            a.ngf * 16,  # encoder_11: [batch, h/32, w/32, ngf * 16] => [batch, h/32, w/32, ngf * 16]
            a.ngf * 16,  # encoder_12: [batch, h/32, w/32, ngf * 16] => [batch, h/64, w/64, ngf * 16]
            a.ngf * 16,  # encoder_13: [batch, h/64, w/64, ngf * 16] => [batch, h/64, w/64, ngf * 16]
        ]
        for out_channels in layer_specs:
            with tf.variable_scope("encoder_%d" % (len(layers) + 1)):
                if (len(layers) + 1) % 2 == 0:
                    # [batch, in_height, in_width, in_channels] => [batch, in_height/2, in_width/2, out_channels]
                    convolved = conv(layers[-1], out_channels, stride=2, shift=4, trainable=trainable)
                else:
                    # [batch, in_height, in_width, in_channels] => [batch, in_height, in_width, out_channels]
                    convolved = conv(layers[-1], out_channels, stride=1, shift=3, trainable=trainable)
                output = batchnorm(convolved, trainable=trainable)
                rectified = tf.nn.relu(output)
                layers.append(rectified)

        # layer_specs = [
        #     a.ngf * 16,  # encoder_10: [batch, h/16, w/16, ngf * 16] => [batch, h/32, w/32, ngf * 16]
        #     a.ngf * 16,  # encoder_11: [batch, h/32, w/32, ngf * 16] => [batch, h/64, w/64, ngf * 16]
        #     a.ngf * 16,  # encoder_12: [batch, h/64, w/64, ngf * 16] => [batch, h/128, w/128, ngf * 16]
        #     a.ngf * 16,  # encoder_13: [batch, h/128, w/128, ngf * 16] => [batch, h/128, w/128, ngf * 16]
        # ]
        #
        # for out_channels_i, out_channels in enumerate(layer_specs):
        #     with tf.variable_scope("encoder_%d" % (len(layers) + 1)):
        #         if out_channels_i != len(layer_specs) -1:
        #             # [batch, in_height, in_width, in_channels] => [batch, in_height/2, in_width/2, out_channels]
        #             convolved = conv(layers[-1], out_channels, stride=2, shift=4, trainable=trainable)
        #         else:
        #             # [batch, in_height, in_width, in_channels] => [batch, in_height/2, in_width/2, out_channels]
        #             convolved = conv(layers[-1], out_channels, stride=1, shift=3, trainable=trainable)
        #         output = batchnorm(convolved, trainable=trainable)
        #         rectified = tf.nn.relu(output)
        #         layers.append(rectified)
        #
        #
        # layer_specs = [
        #     (a.ngf * 16),   # decoder_8: [batch, h/128, w/128, ngf * 16 * 2]=> [batch, h/64, w/64, ngf * 16]
        #     (a.ngf * 16),   # decoder_7: [batch, h/64, w/64, ngf * 16 * 2] => [batch, h/32, w/32, ngf * 16]
        #     (a.ngf * 16),   # decoder_6: [batch, h/32, w/32, ngf * 16 * 2] => [batch, h/16, w/16, ngf * 16]
        # ]
        # num_encoder_layers = len(layers)
        # for decoder_layer, (out_channels) in enumerate(layer_specs):
        #     skip_layer = num_encoder_layers - decoder_layer - 1
        #     with tf.variable_scope("decoder_%d" % (skip_layer + 1)):
        #         if decoder_layer == 0:
        #             input = tf.concat(3, [layers[-1], layers[-2]])
        #         else:
        #             input = tf.concat(3, [layers[-1], layers[skip_layer]])
        #         # [batch, in_height, in_width, in_channels] => [batch, in_height*2, in_width*2, out_channels]
        #         output = deconv(input, out_channels, 2, 4, trainable=trainable)
        #         output = batchnorm(output, trainable=trainable)
        #         output = tf.nn.relu(output)
        #         layers.append(output)

        layer_specs = [
            (a.ngf * 16),  # decoder_12: [batch, h/64, w/64, ngf * 16 * 2]=> [batch, h/32, w/32, ngf * 16]
            (a.ngf * 16),  # decoder_11: [batch, h/32, w/32, ngf * 16]=> [batch, h/32, w/32, ngf * 16]
            (a.ngf * 16),  # decoder_10: [batch, h/32, w/32, ngf * 16 * 2] => [batch, h/16, w/16, ngf * 16]
            (a.ngf * 16),  # decoder_9: [batch, h/16, w/16, ngf * 16] => [batch, h/16, w/16, ngf * 16]
            (a.ngf * 16),  # decoder_8: [batch, h/16, w/16, ngf * 16 * 2]=> [batch, h/8, w/8, ngf * 16]
            (a.ngf * 8),   # decoder_7: [batch, h/8, w/8, ngf * 16] => [batch, h/8, w/8, ngf * 8]
            (a.ngf * 8),   # decoder_6: [batch, h/8, w/8, ngf * 8 * 2] => [batch, h/4, w/4, ngf * 8]
            (a.ngf * 4),   # decoder_5: [batch, h/4, w/4, ngf * 8] => [batch, h/4, w/4, ngf * 4]
            (a.ngf * 4),   # decoder_4: [batch, h/4, w/4, ngf * 4 * 2] => [batch, h/2, w/2, ngf * 4]
            (a.ngf * 2),   # decoder_3: [batch, h/2, w/2, ngf * 4] => [batch, h/2, w/2, ngf * 2]
            (a.ngf * 2),       # decoder_2: [batch, h/2, w/2, ngf * 2 * 2] => [batch, h, w, ngf * 2]
            (a.ngf * 1),                 # decoder_1: [batch, h, w, ngf * 2] => [batch, h, w, ngf]
        ]
        num_encoder_layers = len(layers)
        for decoder_layer, (out_channels) in enumerate(layer_specs):
            skip_layer = num_encoder_layers - decoder_layer - 1
            with tf.variable_scope("decoder_%d" % (skip_layer + 1)):
                if decoder_layer % 2 != 0:
                    # first decoder layer doesn't have skip connections
                    # since it is directly connected to the skip_layer
                    input = layers[-1]
                    # [batch, in_height, in_width, in_channels] => [batch, in_height, in_width, out_channels]
                    output = deconv(input, out_channels, 1, 3, trainable=trainable)
                else:
                    # if decoder_layer == 0:
                    #     input = tf.concat(3, [layers[-1], layers[-2]])
                    # else:
                    #     input = tf.concat(3, [layers[-1], layers[skip_layer]])
                    input = tf.concat(3, [layers[-1], layers[skip_layer]])
                    # [batch, in_height, in_width, in_channels] => [batch, in_height*2, in_width*2, out_channels]
                    output = deconv(input, out_channels, 2, 4, trainable=trainable)
                output = batchnorm(output, trainable=trainable)
                output = tf.nn.relu(output)
                layers.append(output)

        # decoder_1: [batch, h/2, w/2, ngf * 2] => [batch, h, w, generator_outputs_channels]
        with tf.variable_scope("decoder_1"):
            input = tf.concat(3,[layers[-1], layers[0]])
            output = deconv(input, generator_outputs_channels, 1, 3, trainable=trainable)
            layers.append(output)

        return layers[-1]

    def create_discriminator(discrim_inputs, discrim_targets):
        layers = []

        # 2x [batch, height, width, in_channels] => [batch, height, width, in_channels * 2]
        input = tf.concat(3, [discrim_inputs, discrim_targets])

        # layer_1: [batch, h, w, in_channels * 2] => [batch, h/2, w/2, ndf]
        with tf.variable_scope("layer_1"):
            convolved = conv(input, a.ndf, stride=2, shift=4)
            normed = batchnorm(convolved)
            rectified = lrelu(normed, 0.2)
            layers.append(rectified)

        # layer_2: [batch, h/2, w/2, ndf] => [batch, h/4, w/4, ndf * 2]
        # layer_3: [batch, h/4, w/4, ndf * 2] => [batch, h/8, w/8, ndf * 4]
        # layer_4: [batch, h/8, w/8, ndf * 4] => [batch, 31, 31, ndf * 8]
        layer_specs = [
            a.ndf,  # encoder_2: [batch, h, w, ngf] => [batch, h/2, w/2, ngf * 2]
            a.ndf * 2,  # encoder_2: [batch, h, w, ngf] => [batch, h/2, w/2, ngf * 2]
            a.ndf * 2,  # encoder_3: [batch, h/2, w/2, ngf * 2] => [batch, h/2, w/2, ngf * 2]
            a.ndf * 4,  # encoder_4: [batch, h/2, w/2, ngf * 2] => [batch, h/4, w/4, ngf * 4]
            a.ndf * 4,  # encoder_5: [batch, h/4, w/4, ngf * 4] => [batch, h/4, w/4, ngf * 4]
            a.ndf * 8,  # encoder_6: [batch, h/4, w/4, ngf * 4] => [batch, h/8, w/8, ngf * 8]
            a.ndf * 8,  # encoder_7: [batch, h/8, w/8, ngf * 8] => [batch, h/16, w/16, ngf * 8]
            a.ndf * 16,  # encoder_8: [batch, h/16, w/16, ngf * 16] => [batch, h/16, w/16, ngf * 16]
        ]
        for out_channels in layer_specs:
            with tf.variable_scope("layer_%d" % (len(layers) + 1)):
                if (len(layers) + 1) % 2 == 0:
                    # [batch, in_height, in_width, in_channels] => [batch, in_height, in_width, out_channels]
                    convolved = conv(layers[-1], out_channels, stride=1, shift=3)
                else:
                    # [batch, in_height, in_width, in_channels] => [batch, in_height/2, in_width/2, out_channels]
                    convolved = conv(layers[-1], out_channels, stride=2, shift=4)

                normed = batchnorm(convolved)
                rectified = tf.nn.relu(normed)
                layers.append(rectified)

        # layer_5: [batch, h/16, w/16, ngf * 16] => [batch, h/16, w/16, 1]
        with tf.variable_scope("layer_%d" % (len(layers) + 1)):
            # With WGAN, sigmoid for the last layer is no longer needed
            convolved = conv(rectified, out_channels=1, stride=1, shift=3)
            layers.append(convolved)

        return layers[-1]

    if a.use_sketch_loss:
        with tf.variable_scope(SKETCH_VAR_SCOPE_PREFIX + "generator") as scope:
            # real_sketches = create_generator(targets, 1, trainable=False)
            # The commented out piece of code is for using the dilation sketch generation instead of the pretrained
            # sketch generator. TWhen the input image does not contain information about the background, the pretrained
            # sketch generator works better because it ignores the background when generating sketches.
            real_sketches = sketch_extractor(targets, color_space="lab" if a.lab_colorization else "rgb")

    with tf.variable_scope("generator" if not a.train_sketch else SKETCH_VAR_SCOPE_PREFIX + "generator") as scope:
        out_channels = int(targets.get_shape()[-1])
        if a.output_ab:
            assert out_channels == 3
            ab_outputs = create_generator(inputs, 2)
            # Concatenate the l layer in the input with the ab output by the generator.
            outputs = tf.concat(3, (inputs[..., :1], ab_outputs))
            print(outputs.get_shape().as_list())
            assert int(outputs.get_shape()[-1]) == out_channels
        else:
            if a.gen_sketch_input:
                assert a.use_sketch_loss
                # Always use the sketch generated as the input if the gen_sketch_input is on.
                outputs = create_generator(real_sketches, out_channels)
            else:
                outputs = create_generator(inputs,out_channels)

    if a.use_sketch_loss:
        with tf.variable_scope(SKETCH_VAR_SCOPE_PREFIX + "generator", reuse=True) as scope:
            # fake_sketches = create_generator(outputs, 1, trainable=False)
            # Same as the real_sketches line above.
            fake_sketches = sketch_extractor(outputs, color_space="lab" if a.lab_colorization else "rgb")


    # Create two copies of the discriminator: one for real pairs and one for fake pairs.
    # They share the same underlying variables.
    with tf.name_scope("real_discriminator"):
        with tf.variable_scope("discriminator" if not a.train_sketch else SKETCH_VAR_SCOPE_PREFIX + "discriminator"):
            if a.use_hint:
                # Creating discriminator without feeding in the hint information as part of the input. This makes the
                # learning process a little bit easier for the discriminator.
                print(inputs[...,:1].get_shape().as_list())
                predict_real = create_discriminator(inputs[...,:1], targets)
            else:
                predict_real = create_discriminator(inputs, targets)

    with tf.name_scope("fake_discriminator"):
        with tf.variable_scope("discriminator" if not a.train_sketch else SKETCH_VAR_SCOPE_PREFIX + "discriminator", reuse=True):
            if a.use_hint:
                print(inputs[...,:1].get_shape().as_list())
                predict_fake = create_discriminator(inputs[...,:1], outputs)
            else:
                predict_fake = create_discriminator(inputs, outputs)

    if a.use_hint and a.hint_prob > 0:
        is_hint_off_float = tf.cast(is_hint_off, tf.float32, name="is_hint_off_float")

    with tf.name_scope("discriminator_loss"):
        # # minimizing -tf.log will try to get inputs to 1
        # # predict_real => 1
        # # predict_fake => 0
        # discrim_loss = tf.reduce_mean(-(tf.log(predict_real + EPS) + tf.log(1 - predict_fake + EPS)))

        # # Use wgan loss
        # discrim_loss = -tf.reduce_mean(predict_real) + tf.reduce_mean(predict_fake)

        # Use wgan loss, turn on loss only when hint is off.
        # I have to do the div in each item to be compatible with the previous version's checkpoints.
        if a.use_hint and a.hint_prob >0:
            discrim_loss = -tf.div(tf.reduce_mean(predict_real * is_hint_off_float), min(a.hint_prob, 1)) + tf.div(tf.reduce_mean(predict_fake * is_hint_off_float), min(a.hint_prob, 1))
        else:
            discrim_loss = -tf.reduce_mean(predict_real) + tf.reduce_mean(predict_fake)

    with tf.name_scope("generator_loss"):
        # # predict_fake => 1
        # # abs(targets - outputs) => 0
        # gen_loss_GAN = tf.reduce_mean(-tf.log(predict_fake + EPS))

        # # WGAN loss
        # gen_loss_GAN = -tf.reduce_mean(predict_fake)

        # WGAN loss, turn on loss only when hint is off.
        if a.use_hint and a.hint_prob >0:
            gen_loss_GAN = -tf.div(tf.reduce_mean(predict_fake * is_hint_off_float), min(a.hint_prob, 1))
        else:
            gen_loss_GAN = -tf.reduce_mean(predict_fake)

        gen_loss_L1 = tf.reduce_mean(tf.abs(targets - outputs))
        if a.use_sketch_loss:
            gen_loss_sketch = tf.reduce_mean(tf.abs(fake_sketches - real_sketches))
            gen_loss = gen_loss_GAN * a.gan_weight + gen_loss_L1 * a.l1_weight + gen_loss_sketch * a.sketch_weight
        else:
            gen_loss = gen_loss_GAN * a.gan_weight + gen_loss_L1 * a.l1_weight

    with tf.name_scope("discriminator_train"):
        discrim_tvars = [var for var in tf.trainable_variables() if var.name.startswith("discriminator" if not a.train_sketch else SKETCH_VAR_SCOPE_PREFIX + "discriminator")]
        # discrim_optim = tf.train.AdamOptimizer(a.lr, a.beta1)
        # WGAN does not use momentum based optimizer
        discrim_optim = tf.train.RMSPropOptimizer(a.lr)
        # discrim_train = discrim_optim.minimize(discrim_loss, var_list=discrim_tvars)

        # WGAN adds a clip and train discriminator 5 times
        discrim_min = discrim_optim.minimize(discrim_loss, var_list=discrim_tvars)
        discrim_clips = [var.assign(tf.clip_by_value(var, -CLIP_VALUE, CLIP_VALUE)) for var in discrim_tvars]
        with tf.control_dependencies([discrim_min]):
            discrim_train = tf.group(*discrim_clips)

    with tf.name_scope("generator_train"):
        with tf.control_dependencies([discrim_train]):
            gen_tvars = [var for var in tf.trainable_variables() if var.name.startswith("generator" if not a.train_sketch else SKETCH_VAR_SCOPE_PREFIX + "generator")]
            # gen_optim = tf.train.AdamOptimizer(a.lr, a.beta1)
            gen_optim = tf.train.RMSPropOptimizer(a.lr)
            gen_train = gen_optim.minimize(gen_loss, var_list=gen_tvars)

    ema = tf.train.ExponentialMovingAverage(decay=0.99)
    if a.use_sketch_loss:
        update_losses = ema.apply([discrim_loss, gen_loss_GAN, gen_loss_L1, gen_loss_sketch])
    else:
        update_losses = ema.apply([discrim_loss, gen_loss_GAN, gen_loss_L1])

    global_step = tf.contrib.framework.get_or_create_global_step()
    incr_global_step = tf.assign(global_step, global_step+1)

    return Model(
        predict_real=predict_real,
        predict_fake=predict_fake,
        real_sketches=real_sketches if a.use_sketch_loss else None,
        fake_sketches=fake_sketches if a.use_sketch_loss else None,
        discrim_loss=ema.average(discrim_loss),
        gen_loss_GAN=ema.average(gen_loss_GAN),
        gen_loss_L1=ema.average(gen_loss_L1),
        gen_loss_sketch=ema.average(gen_loss_sketch) if a.use_sketch_loss else None,
        outputs=outputs,
        train=tf.group(update_losses, incr_global_step, gen_train),
    )


def save_images(fetches, image_dir, step=None):
    filesets = []
    for i, in_path in enumerate(fetches["paths"]):
        name, _ = os.path.splitext(os.path.basename(in_path))
        fileset = {"name": name, "step": step}
        kinds = ["outputs", ]
        if not (a.single_input and a.mode == "test"):
            kinds = kinds + ["inputs", "targets"]
            if a.use_hint:
                kinds.append("hints")
            if a.use_sketch_loss:
                kinds = kinds + ["real_sketches", "fake_sketches"]
        for kind in kinds:
            if (a.single_input and a.mode == "test"):
                # Do not modify file name when single input.
                filename = name + ".png"
            else:
                filename = name + "-" + kind + ".png"
                if step is not None:
                    filename = "%08d-%s" % (step, filename)
            fileset[kind] = filename
            out_path = os.path.join(image_dir, filename)
            contents = fetches[kind][i]
            with open(out_path, "w") as f:
                f.write(contents)
        filesets.append(fileset)
    return filesets


def append_index(filesets, step=False):
    index_path = os.path.join(a.output_dir, "index.html")
    if os.path.exists(index_path):
        index = open(index_path, "a")
    else:
        index = open(index_path, "w")
        index.write("<html><meta content=\"text/html;charset=utf-8\" http-equiv=\"Content-Type\"><meta content=\"utf-8\" http-equiv=\"encoding\"><body><table><tr>")
        if step:
            index.write("<th>step</th>")
        if not (a.single_input and a.mode == "test"):
            if a.use_hint:
                if a.use_sketch_loss:
                    index.write("<th>name</th><th>input</th><th>hint</th><th>output</th><th>target</th><th>real_sketch</th><th>fake_sketch</th></tr>")
                else:
                    index.write("<th>name</th><th>input</th><th>hint</th><th>output</th><th>target</th></tr>")
            else:
                if a.use_sketch_loss:
                    index.write("<th>name</th><th>input</th><th>output</th><th>target</th><th>real_sketch</th><th>fake_sketch</th></tr>")
                else:
                    index.write("<th>name</th><th>input</th><th>output</th><th>target</th></tr>")
        else:
            index.write("<th>name</th><th>output</th></tr>")

    for fileset in filesets:
        index.write("<tr>")
        if step:
            index.write("<td>%d</td>" % fileset["step"])
        index.write("<td>%s</td>" % fileset["name"])
        kinds = ["outputs", ]
        if not (a.single_input and a.mode == "test"):
            kinds = ["inputs", "hints", "outputs", "targets"] if a.use_hint else ["inputs", "outputs", "targets"]
            if a.use_sketch_loss:
                kinds = kinds + ["real_sketches", "fake_sketches"]
        for kind in kinds:
            index.write("<td><img src=\"images/%s\"></td>" % urllib.quote(fileset[kind]))

        index.write("</tr>")
    return index_path

def main():
    if a.from_128:
        assert a.checkpoint is not None

    if a.seed is None:
        a.seed = random.randint(0, 2**31 - 1)

    tf.set_random_seed(a.seed)
    np.random.seed(a.seed)
    random.seed(a.seed)

    if not os.path.exists(a.output_dir):
        os.makedirs(a.output_dir)

    if a.mode == "test":
        if a.checkpoint is None:
            raise Exception("checkpoint required for test mode")

        # load some options from the checkpoint
        options = {"which_direction", "ngf", "ndf", "lab_colorization", "gray_input_a", "gray_input_b", "use_hint"}
        with open(os.path.join(a.checkpoint, "options.json")) as f:
            for key, val in json.loads(f.read()).iteritems():
                if key in options:
                    print("loaded", key, "=", val)
                    setattr(a, key, val)
        # disable these features in test mode
        a.scale_size = CROP_SIZE
        a.flip = False
        if not (a.mix_prob == 1.0 or a.mix_prob <= 0):
            input("Are you sure you don't want to set the mix_prob to 1 (always using the simple sketch generator) or "
                  "0 (always use the input image)?")
        if a.user_hint_path is not None:
            a.hint_prob = 1.0
            if a.user_hint_path == "BLANK":
                user_hint = np.ones([CROP_SIZE,CROP_SIZE,4], dtype=np.float32) * -1
            else:
                user_hint = (np.array(imread(a.user_hint_path, shape=(CROP_SIZE,CROP_SIZE), rgba=True)) - 255.0/2) / (255.0/2)
                assert np.max(user_hint) <=1.0 and np.min(user_hint) >=-1.0
        else:
            user_hint = None

    else:
        assert a.user_hint_path == None  # Can't train with one single specified hint image. Doesn't make sense.
        user_hint = None

    for k, v in a._get_kwargs():
        print(k, "=", v)

    with open(os.path.join(a.output_dir, "options.json"), "w") as f:
        f.write(json.dumps(vars(a), sort_keys=True, indent=4))

    examples = load_examples(user_hint=user_hint)

    print("examples count = %d" % examples.count)

    model = create_model(examples.inputs, examples.targets, examples.is_hint_off)

    def deprocess(image):
        if a.aspect_ratio != 1.0:
            # upscale to correct aspect ratio
            size = [CROP_SIZE, int(round(CROP_SIZE * a.aspect_ratio))]
            image = tf.image.resize_images(image, size=size, method=tf.image.ResizeMethod.BICUBIC)

        if a.lab_colorization:
            # colorization mode images can be 1 channel (L) or 2 channels (a,b)
            num_channels = int(image.get_shape()[-1])
            if num_channels == 1:
                return tf.image.convert_image_dtype((image + 1) / 2, dtype=tf.uint8, saturate=True)
            elif num_channels == 3:
                # (a, b) color channels, convert to rgb
                # a_chan and b_chan have range [-1, 1] => [-110, 110]
                L_chan, a_chan, b_chan = tf.unstack(image, axis=3)
                L_chan = (L_chan + 1) * 50
                a_chan = a_chan * 110
                b_chan = b_chan * 110
                lab = tf.stack([L_chan, a_chan, b_chan], axis=3)
                rgb = lab_to_rgb(lab)
                return tf.image.convert_image_dtype(rgb, dtype=tf.uint8, saturate=True)
            elif num_channels == 4:
                print('using hint! correct!')
                # (a, b) color channels, convert to rgb
                # a_chan and b_chan have range [-1, 1] => [-110, 110]
                L_chan, a_chan, b_chan, opacity_chan = tf.unstack(image, axis=3)
                L_chan = (L_chan + 1) * 50
                a_chan = a_chan * 110
                b_chan = b_chan * 110
                lab = tf.stack([L_chan, a_chan, b_chan], axis=3)
                rgb = lab_to_rgb(lab)
                opacity_chan = tf.expand_dims((opacity_chan + 1) / 2,3)  # [-1, 1] => [0, 1]
                rgba = tf.concat(3, (rgb, opacity_chan))
                return tf.image.convert_image_dtype(rgba, dtype=tf.uint8, saturate=True)
            else:
                raise Exception("unexpected number of channels")
        else:
            num_channels = int(image.get_shape()[-1])
            if num_channels != 4:
                return tf.image.convert_image_dtype((image + 1) / 2, dtype=tf.uint8, saturate=True)
            else:
                print('using hint! correct!')
                print('image shape: %s' %(str(image.get_shape().as_list())))
                # image_r,image_g,image_b,image_a = tf.unpack(image,axis=3)
                # image_r = (image_r + 1) / 2
                # image_g = (image_g + 1) / 2
                # image_b = (image_b + 1) / 2
                # image_rgba = tf.pack((image_r,image_g,image_b,image_a),axis=3)
                image_rgba = (image + 1) / 2
                return tf.image.convert_image_dtype(image_rgba, dtype=tf.uint8, saturate=True)

    # reverse any processing on images so they can be written to disk or displayed to user
    with tf.name_scope("deprocess_inputs"):
        if a.use_hint:
            deprocessed_inputs = deprocess(examples.inputs[...,:1])
            deprocessed_hints = deprocess(examples.inputs[...,1:])
        else:
            deprocessed_inputs = deprocess(examples.inputs)


    with tf.name_scope("deprocess_targets"):
        deprocessed_targets = deprocess(examples.targets)

    with tf.name_scope("deprocess_outputs"):
        deprocessed_outputs = deprocess(model.outputs)

    if a.use_sketch_loss:
        with tf.name_scope("deprocess_real_sketches"):
            deprocessed_real_sketches = deprocess(model.real_sketches)
        with tf.name_scope("deprocess_fake_sketches"):
            deprocessed_fake_sketches = deprocess(model.fake_sketches)

    with tf.name_scope("encode_images"):
        if a.use_hint:
            if a.use_sketch_loss:
                display_fetches = {
                    "paths": examples.paths,
                    "inputs": tf.map_fn(tf.image.encode_png, deprocessed_inputs, dtype=tf.string, name="input_pngs"),
                    "hints": tf.map_fn(tf.image.encode_png, deprocessed_hints, dtype=tf.string, name="hint_pngs"),
                    "targets": tf.map_fn(tf.image.encode_png, deprocessed_targets, dtype=tf.string, name="target_pngs"),
                    "outputs": tf.map_fn(tf.image.encode_png, deprocessed_outputs, dtype=tf.string, name="output_pngs"),
                    "real_sketches": tf.map_fn(tf.image.encode_png, deprocessed_real_sketches, dtype=tf.string, name="real_sketches_pngs"),
                    "fake_sketches": tf.map_fn(tf.image.encode_png, deprocessed_fake_sketches, dtype=tf.string, name="fake_sketches_pngs"),
                }
            else:
                display_fetches = {
                    "paths": examples.paths,
                    "inputs": tf.map_fn(tf.image.encode_png, deprocessed_inputs, dtype=tf.string, name="input_pngs"),
                    "hints": tf.map_fn(tf.image.encode_png, deprocessed_hints, dtype=tf.string, name="hint_pngs"),
                    "targets": tf.map_fn(tf.image.encode_png, deprocessed_targets, dtype=tf.string, name="target_pngs"),
                    "outputs": tf.map_fn(tf.image.encode_png, deprocessed_outputs, dtype=tf.string, name="output_pngs"),
                }
        else:
            if a.use_sketch_loss:
                display_fetches = {
                    "paths": examples.paths,
                    "inputs": tf.map_fn(tf.image.encode_png, deprocessed_inputs, dtype=tf.string, name="input_pngs"),
                    "targets": tf.map_fn(tf.image.encode_png, deprocessed_targets, dtype=tf.string, name="target_pngs"),
                    "outputs": tf.map_fn(tf.image.encode_png, deprocessed_outputs, dtype=tf.string, name="output_pngs"),
                    "real_sketches": tf.map_fn(tf.image.encode_png, deprocessed_real_sketches, dtype=tf.string, name="real_sketches_pngs"),
                    "fake_sketches": tf.map_fn(tf.image.encode_png, deprocessed_fake_sketches, dtype=tf.string, name="fake_sketches_pngs"),
                }
            else:
                display_fetches = {
                    "paths": examples.paths,
                    "inputs": tf.map_fn(tf.image.encode_png, deprocessed_inputs, dtype=tf.string, name="input_pngs"),
                    "targets": tf.map_fn(tf.image.encode_png, deprocessed_targets, dtype=tf.string, name="target_pngs"),
                    "outputs": tf.map_fn(tf.image.encode_png, deprocessed_outputs, dtype=tf.string, name="output_pngs"),
                }


    # summaries
    with tf.name_scope("inputs_summary"):
        tf.summary.image("inputs", deprocessed_inputs)
    if a.use_hint:
        with tf.name_scope("hints_summary"):
            tf.summary.image("inputs", deprocessed_hints)

    with tf.name_scope("targets_summary"):
        tf.summary.image("targets", deprocessed_targets)

    with tf.name_scope("outputs_summary"):
        tf.summary.image("outputs", deprocessed_outputs)

    if a.use_sketch_loss:
        with tf.name_scope("real_sketches_summary"):
            tf.summary.image("real_sketches", deprocessed_real_sketches)
        with tf.name_scope("fake_sketches_summary"):
            tf.summary.image("fake_sketches", deprocessed_fake_sketches)

    with tf.name_scope("predict_real_summary"):
        # In order for the image to be within range 0~1, I need to apply tanh here.
        tf.summary.image("predict_real", tf.image.convert_image_dtype((tf.nn.tanh(model.predict_real) + 1) / 2, dtype=tf.uint8))

    with tf.name_scope("predict_fake_summary"):
        tf.summary.image("predict_fake", tf.image.convert_image_dtype((tf.nn.tanh(model.predict_fake)) + 1 / 2, dtype=tf.uint8))

    tf.summary.scalar("discriminator_loss", model.discrim_loss)
    tf.summary.scalar("generator_loss_GAN", model.gen_loss_GAN)
    tf.summary.scalar("generator_loss_L1", model.gen_loss_L1)
    if a.use_sketch_loss:
        tf.summary.scalar("gen_loss_sketch", model.gen_loss_sketch)

    with tf.name_scope("parameter_count"):
        parameter_count = tf.reduce_sum([tf.reduce_prod(tf.shape(v)) for v in tf.trainable_variables()])

    image_dir = os.path.join(a.output_dir, "images")
    if not os.path.exists(image_dir):
        os.makedirs(image_dir)

    config = tf.ConfigProto()
    if a.gpu_percentage > 0:
        config.gpu_options.per_process_gpu_memory_fraction = a.gpu_percentage
    else:
        config = tf.ConfigProto(
            device_count = {'GPU': 0}
        )

    if a.from_128:
        generator_var = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope='generator' if not a.train_sketch else SKETCH_VAR_SCOPE_PREFIX + "generator")
        discriminator_var = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope='discriminator' if not a.train_sketch else SKETCH_VAR_SCOPE_PREFIX + "discriminator")
        other_var = [var for var in tf.global_variables() if (var not in generator_var and var not in discriminator_var)]
        print("number of generator var = %d, number of discriminator var = %d, number of other var = %d"
              %(len(generator_var), len(discriminator_var), len(other_var)))
        assert len(generator_var) != 0 and len(discriminator_var) != 0
        saver = tf.train.Saver(max_to_keep=1, var_list=generator_var + discriminator_var)
        with tf.Session(config=config) as sess:
            print("loading model from checkpoint")
            checkpoint = tf.train.latest_checkpoint(a.checkpoint)
            saver.restore(sess, checkpoint)
            sess.run(tf.initialize_variables(other_var))
            saver = tf.train.Saver(max_to_keep=1)
            saver.save(sess,checkpoint)
    else:
        # If there is a checkpoint, then the sketch generator variables should already be stored in there.
        # if a.use_sketch_loss and a.mode != "test" and a.checkpoint is None:
        #     sketch_var = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope=SKETCH_VAR_SCOPE_PREFIX + "generator")
        #     # This is a sanity check to make sure sketch variables are not trainable.
        #     assert len(tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,
        #                                    scope=SKETCH_VAR_SCOPE_PREFIX + "generator")) == 0
        #     other_var = [var for var in tf.global_variables() if
        #                  (var not in sketch_var)]
        #     print("number of sketch var = %d, number of other var = %d"
        #           % (len(sketch_var), len(other_var)))
        #     assert len(sketch_var) != 0
        #     saver = tf.train.Saver(max_to_keep=1, var_list=sketch_var)
        #     with tf.Session(config=config) as sess:
        #         print("loading sketch generator model from checkpoint")
        #         pretrained_sketch_checkpoint = tf.train.latest_checkpoint(a.pretrained_sketch_net_path)
        #         saver.restore(sess, pretrained_sketch_checkpoint)
        #         sess.run(tf.initialize_variables(other_var))
        #         sketch_var_value_before = sketch_var[0].eval()[0, 0, 0, 0]
        #         print("Sketch var value before supervised session: %f" %(sketch_var_value_before))
        #         saver = tf.train.Saver(max_to_keep=1)
        #         saver.save(sess,save_path=os.path.join(a.output_dir, "model"))
        # else:
        #     saver = tf.train.Saver(max_to_keep=1)
        # The code commented out is for when one wants to switch from the pretrained sketch generator for sketch loss
        # to the dilation sketch generator.
        sketch_var = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope=SKETCH_VAR_SCOPE_PREFIX + "generator")
        if not a.train_sketch:
            assert len(sketch_var) == 0
        saver = tf.train.Saver(max_to_keep=1)

    logdir = a.output_dir if (a.trace_freq > 0 or a.summary_freq > 0) else None
    sv = tf.train.Supervisor(logdir=logdir, save_summaries_secs=0, saver=None)
    with sv.managed_session(config=config) as sess:
        print("parameter_count =", sess.run(parameter_count))

        if a.checkpoint is not None:
            print("loading model from checkpoint")
            checkpoint = tf.train.latest_checkpoint(a.checkpoint)
            saver.restore(sess, checkpoint)
        # elif a.use_sketch_loss:
        #     print("loading model with saved sketch generator variables.")
        #     checkpoint = tf.train.latest_checkpoint(a.output_dir)
        #     print(checkpoint)
        #     assert checkpoint is not None
        #     saver.restore(sess, checkpoint)
        #
        #     # sketch_var_value_after = sketch_var[0].eval(session=sess)[0, 0, 0, 0]
        #     # print("Sketch var value after supervised session: %f" %(sketch_var_value_after))
        #     # assert sketch_var_value_after == sketch_var_value_before



        if a.mode == "test":
            # Testing, run a single epoch over all input data
            for step in range(examples.steps_per_epoch):
                results = sess.run(display_fetches)
                filesets = save_images(results, image_dir)
                if not a.single_input:
                    for i, path in enumerate(results["paths"]):
                        print(step * a.batch_size + i + 1, "evaluated image", os.path.basename(path))
                else:
                    if step % 100 == 0:
                        print("Evaluated %d out of %d steps." %(step,examples.steps_per_epoch))
                index_path = append_index(filesets)
            print("wrote index at", index_path)
        else:
            # Training
            max_steps = 2**32
            if a.max_epochs is not None:
                max_steps = examples.steps_per_epoch * a.max_epochs
            if a.max_steps is not None:
                max_steps = a.max_steps

            initial_global_step, = sess.run([sv.global_step])

            start_time = time.time()
            for step in range(max_steps):
                def should(freq):
                    return freq > 0 and ((step + 1) % freq == 0 or step == max_steps - 1)

                options = None
                run_metadata = None
                if should(a.trace_freq):
                    options = tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE)
                    run_metadata = tf.RunMetadata()

                fetches = {
                    "train": model.train,
                    "global_step": sv.global_step,
                }

                if should(a.progress_freq):
                    fetches["discrim_loss"] = model.discrim_loss
                    fetches["gen_loss_GAN"] = model.gen_loss_GAN
                    fetches["gen_loss_L1"] = model.gen_loss_L1
                    if a.use_sketch_loss:
                        fetches["gen_loss_sketch"] = model.gen_loss_sketch

                if should(a.summary_freq):
                    fetches["summary"] = sv.summary_op

                if should(a.display_freq):
                    fetches["display"] = display_fetches

                results = sess.run(fetches, options=options, run_metadata=run_metadata)

                if should(a.summary_freq):
                    sv.summary_writer.add_summary(results["summary"], results["global_step"])

                if should(a.display_freq):
                    print("saving display images")
                    filesets = save_images(results["display"], image_dir, step=results["global_step"])
                    append_index(filesets, step=True)

                if should(a.trace_freq):
                    print("recording trace")
                    sv.summary_writer.add_run_metadata(run_metadata, "step_%d" % results["global_step"])

                if should(a.progress_freq):
                    global_step = results["global_step"]
                    print("progress  epoch %d  step %d  image/sec %0.1f" % (global_step // examples.steps_per_epoch, global_step % examples.steps_per_epoch, (global_step - initial_global_step) * a.batch_size / (time.time() - start_time)))
                    print("discrim_loss", results["discrim_loss"])
                    print("gen_loss_GAN", results["gen_loss_GAN"])
                    print("gen_loss_L1", results["gen_loss_L1"])
                    if a.use_sketch_loss:
                        print("gen_loss_sketch", results["gen_loss_sketch"])

                if should(a.save_freq):
                    print("saving model")
                    saver.save(sess, os.path.join(a.output_dir, "model"), global_step=sv.global_step)

                if sv.should_stop():
                    break

if __name__ == "__main__":
    main()