stylize.py

# Copyright (c) Anish Athalye. Released under GPLv3.

import os
import time
from collections import OrderedDict

from PIL import Image
import numpy as np
import tensorflow.compat.v1 as tf

tf.disable_v2_behavior()

import vgg


CONTENT_LAYERS = ("relu4_2", "relu5_2")
STYLE_LAYERS = ("relu1_1", "relu2_1", "relu3_1", "relu4_1", "relu5_1")


try:
    reduce
except NameError:
    from functools import reduce


def get_loss_vals(loss_store):
    return OrderedDict((key, val.eval()) for key, val in loss_store.items())


def print_progress(loss_vals):
    for key, val in loss_vals.items():
        print("{:>13s} {:g}".format(key + " loss:", val))


def stylize(
    network,
    initial,
    initial_noiseblend,
    content,
    styles,
    preserve_colors,
    iterations,
    content_weight,
    content_weight_blend,
    style_weight,
    style_layer_weight_exp,
    style_blend_weights,
    tv_weight,
    learning_rate,
    beta1,
    beta2,
    epsilon,
    pooling,
    print_iterations=None,
    checkpoint_iterations=None,
):
    """
    Stylize images.

    This function yields tuples (iteration, image, loss_vals) at every
    iteration. However `image` and `loss_vals` are None by default. Each
    `checkpoint_iterations`, `image` is not None. Each `print_iterations`,
    `loss_vals` is not None.

    `loss_vals` is a dict with loss values for the current iteration, e.g.
    ``{'content': 1.23, 'style': 4.56, 'tv': 7.89, 'total': 13.68}``.

    :rtype: iterator[tuple[int,image]]
    """
    shape = (1,) + content.shape
    style_shapes = [(1,) + style.shape for style in styles]
    content_features = {}
    style_features = [{} for _ in styles]

    vgg_weights, vgg_mean_pixel = vgg.load_net(network)

    layer_weight = 1.0
    style_layers_weights = {}
    for style_layer in STYLE_LAYERS:
        style_layers_weights[style_layer] = layer_weight
        layer_weight *= style_layer_weight_exp

    # normalize style layer weights
    layer_weights_sum = 0
    for style_layer in STYLE_LAYERS:
        layer_weights_sum += style_layers_weights[style_layer]
    for style_layer in STYLE_LAYERS:
        style_layers_weights[style_layer] /= layer_weights_sum

    # compute content features in feedforward mode
    g = tf.Graph()
    with g.as_default(), g.device("/cpu:0"), tf.Session() as sess:
        image = tf.placeholder("float", shape=shape)
        net = vgg.net_preloaded(vgg_weights, image, pooling)
        content_pre = np.array([vgg.preprocess(content, vgg_mean_pixel)])
        for layer in CONTENT_LAYERS:
            content_features[layer] = net[layer].eval(feed_dict={image: content_pre})

    # compute style features in feedforward mode
    for i in range(len(styles)):
        g = tf.Graph()
        with g.as_default(), g.device("/cpu:0"), tf.Session() as sess:
            image = tf.placeholder("float", shape=style_shapes[i])
            net = vgg.net_preloaded(vgg_weights, image, pooling)
            style_pre = np.array([vgg.preprocess(styles[i], vgg_mean_pixel)])
            for layer in STYLE_LAYERS:
                features = net[layer].eval(feed_dict={image: style_pre})
                features = np.reshape(features, (-1, features.shape[3]))
                gram = np.matmul(features.T, features) / features.size
                style_features[i][layer] = gram

    initial_content_noise_coeff = 1.0 - initial_noiseblend

    # make stylized image using backpropogation
    with tf.Graph().as_default():
        if initial is None:
            noise = np.random.normal(size=shape, scale=np.std(content) * 0.1)
            initial = tf.random_normal(shape) * 0.256
        else:
            initial = np.array([vgg.preprocess(initial, vgg_mean_pixel)])
            initial = initial.astype("float32")
            noise = np.random.normal(size=shape, scale=np.std(content) * 0.1)
            initial = (initial) * initial_content_noise_coeff + (
                tf.random_normal(shape) * 0.256
            ) * (1.0 - initial_content_noise_coeff)
        image = tf.Variable(initial)
        net = vgg.net_preloaded(vgg_weights, image, pooling)

        # content loss
        content_layers_weights = {}
        content_layers_weights["relu4_2"] = content_weight_blend
        content_layers_weights["relu5_2"] = 1.0 - content_weight_blend

        content_loss = 0
        content_losses = []
        for content_layer in CONTENT_LAYERS:
            content_losses.append(
                content_layers_weights[content_layer]
                * content_weight
                * (
                    2
                    * tf.nn.l2_loss(net[content_layer] - content_features[content_layer])
                    / content_features[content_layer].size
                )
            )
        content_loss += reduce(tf.add, content_losses)

        # style loss
        style_loss = 0
        for i in range(len(styles)):
            style_losses = []
            for style_layer in STYLE_LAYERS:
                layer = net[style_layer]
                _, height, width, number = map(lambda i: i.value, layer.get_shape())
                size = height * width * number
                feats = tf.reshape(layer, (-1, number))
                gram = tf.matmul(tf.transpose(feats), feats) / size
                style_gram = style_features[i][style_layer]
                style_losses.append(
                    style_layers_weights[style_layer]
                    * 2
                    * tf.nn.l2_loss(gram - style_gram)
                    / style_gram.size
                )
            style_loss += style_weight * style_blend_weights[i] * reduce(tf.add, style_losses)

        # total variation denoising
        tv_y_size = _tensor_size(image[:, 1:, :, :])
        tv_x_size = _tensor_size(image[:, :, 1:, :])
        tv_loss = (
            tv_weight
            * 2
            * (
                (tf.nn.l2_loss(image[:, 1:, :, :] - image[:, : shape[1] - 1, :, :]) / tv_y_size)
                + (tf.nn.l2_loss(image[:, :, 1:, :] - image[:, :, : shape[2] - 1, :]) / tv_x_size)
            )
        )

        # total loss
        loss = content_loss + style_loss + tv_loss

        # We use OrderedDict to make sure we have the same order of loss types
        # (content, tv, style, total) as defined by the initial costruction of
        # the loss_store dict. This is important for print_progress() and
        # saving loss_arrs (column order) in the main script.
        #
        # Subtle Gotcha (tested with Python 3.5): The syntax
        # OrderedDict(key1=val1, key2=val2, ...) does /not/ create the same
        # order since, apparently, it first creates a normal dict with random
        # order (< Python 3.7) and then wraps that in an OrderedDict. We have
        # to pass in a data structure which is already ordered. I'd call this a
        # bug, since both constructor syntax variants result in different
        # objects. In 3.6, the order is preserved in dict() in CPython, in 3.7
        # they finally made it part of the language spec. Thank you!
        loss_store = OrderedDict(
            [("content", content_loss), ("style", style_loss), ("tv", tv_loss), ("total", loss)]
        )

        # optimizer setup
        train_step = tf.train.AdamOptimizer(learning_rate, beta1, beta2, epsilon).minimize(loss)

        # optimization
        best_loss = float("inf")
        best = None
        with tf.Session() as sess:
            sess.run(tf.global_variables_initializer())
            print("Optimization started...")
            if print_iterations and print_iterations != 0:
                print_progress(get_loss_vals(loss_store))
            iteration_times = []
            start = time.time()
            for i in range(iterations):
                iteration_start = time.time()
                if i > 0:
                    elapsed = time.time() - start
                    # take average of last couple steps to get time per iteration
                    remaining = np.mean(iteration_times[-10:]) * (iterations - i)
                    print(
                        "Iteration %4d/%4d (%s elapsed, %s remaining)"
                        % (i + 1, iterations, hms(elapsed), hms(remaining))
                    )
                else:
                    print("Iteration %4d/%4d" % (i + 1, iterations))
                train_step.run()

                last_step = i == iterations - 1
                if last_step or (print_iterations and i % print_iterations == 0):
                    loss_vals = get_loss_vals(loss_store)
                    print_progress(loss_vals)
                else:
                    loss_vals = None

                if (checkpoint_iterations and i % checkpoint_iterations == 0) or last_step:
                    this_loss = loss.eval()
                    if this_loss < best_loss:
                        best_loss = this_loss
                        best = image.eval()

                    img_out = vgg.unprocess(best.reshape(shape[1:]), vgg_mean_pixel)

                    if preserve_colors:
                        original_image = np.clip(content, 0, 255)
                        styled_image = np.clip(img_out, 0, 255)

                        # Luminosity transfer steps:
                        # 1. Convert stylized RGB->grayscale accoriding to Rec.601 luma (0.299, 0.587, 0.114)
                        # 2. Convert stylized grayscale into YUV (YCbCr)
                        # 3. Convert original image into YUV (YCbCr)
                        # 4. Recombine (stylizedYUV.Y, originalYUV.U, originalYUV.V)
                        # 5. Convert recombined image from YUV back to RGB

                        # 1
                        styled_grayscale = rgb2gray(styled_image)
                        styled_grayscale_rgb = gray2rgb(styled_grayscale)

                        # 2
                        styled_grayscale_yuv = np.array(
                            Image.fromarray(styled_grayscale_rgb.astype(np.uint8)).convert("YCbCr")
                        )

                        # 3
                        original_yuv = np.array(
                            Image.fromarray(original_image.astype(np.uint8)).convert("YCbCr")
                        )

                        # 4
                        w, h, _ = original_image.shape
                        combined_yuv = np.empty((w, h, 3), dtype=np.uint8)
                        combined_yuv[..., 0] = styled_grayscale_yuv[..., 0]
                        combined_yuv[..., 1] = original_yuv[..., 1]
                        combined_yuv[..., 2] = original_yuv[..., 2]

                        # 5
                        img_out = np.array(Image.fromarray(combined_yuv, "YCbCr").convert("RGB"))
                else:
                    img_out = None

                yield i + 1 if last_step else i, img_out, loss_vals

                iteration_end = time.time()
                iteration_times.append(iteration_end - iteration_start)


def _tensor_size(tensor):
    from operator import mul

    return reduce(mul, (d.value for d in tensor.get_shape()), 1)


def rgb2gray(rgb):
    return np.dot(rgb[..., :3], [0.299, 0.587, 0.114])


def gray2rgb(gray):
    w, h = gray.shape
    rgb = np.empty((w, h, 3), dtype=np.float32)
    rgb[:, :, 2] = rgb[:, :, 1] = rgb[:, :, 0] = gray
    return rgb


def hms(seconds):
    seconds = int(seconds)
    hours = seconds // (60 * 60)
    minutes = (seconds // 60) % 60
    seconds = seconds % 60
    if hours > 0:
        return "%d hr %d min" % (hours, minutes)
    elif minutes > 0:
        return "%d min %d sec" % (minutes, seconds)
    else:
        return "%d sec" % seconds