Github - justinpinkney/stylegan2

基于 StyleGAN2 的人脸艺术化，项目效果如：

这里主要记录下测试过程.

1. 配置

下载项目:

git clone https://github.com/justinpinkney/stylegan2
cd stylegan2
nvcc test_nvcc.cu -o test_nvcc -run
#tensorflow 1.x

下载权重文件：

#ffhq-cartoon-blended-64.pkl
blended_url = "https://drive.google.com/uc?id=1H73TfV5gQ9ot7slSed_l-lim9X7pMRiU" 

#stylegan2-ffhq-config-f.pkl
ffhq_url = "http://d36zk2xti64re0.cloudfront.net/stylegan2/networks/stylegan2-ffhq-config-f.pkl"

#如：
_, _, Gs_blended = pretrained_networks.load_networks(blended_url)
_, _, Gs = pretrained_networks.load_networks(ffhq_url)

下载测试图片，假设放置路径为 raw/. 建议测试图像分辨率高一些.

wget https://upload.wikimedia.org/wikipedia/commons/6/6d/Shinz%C5%8D_Abe_Official.jpg -O raw/example.jpg

2. 测试

对于待测试图片，测试流程主要为：

  • 人脸检测和对齐
  • 人脸投影(人脸编码，如，计算人脸的 latent code)
  • 人脸卡通化(如，将 latent code 应用于卡通模型)

2.1. 人脸对齐

Dlib库 - 人脸关键点检测与对齐实现 - AIUAI

人脸对齐实现：https://github.com/justinpinkney/stylegan2/blob/master/align_images.py

运行如下，假设对齐后的路径为 aligned/：

python align_images.py raw/ aligned/

对齐后效果如图：

2.2. 人脸投影

运行，假设生成路径为 generated/，

python project_images.py --num-steps 500 aligned/ generated/

2.2.1. projector

projector.py

import numpy as np
import tensorflow as tf
import dnnlib
import dnnlib.tflib as tflib

from training import misc

#----------------------------------------------------------------------------

class Projector:
    def __init__(self,
        vgg16_pkl                       = 'https://drive.google.com/uc?id=1N2-m9qszOeVC9Tq77WxsLnuWwOedQiD2',
        num_steps                       = 1000,
        initial_learning_rate           = 0.1,
        initial_noise_factor            = 0.05,
        verbose                         = False
    ):

        self.vgg16_pkl                  = vgg16_pkl
        self.num_steps                  = num_steps
        self.dlatent_avg_samples        = 10000
        self.initial_learning_rate      = initial_learning_rate
        self.initial_noise_factor       = initial_noise_factor
        self.lr_rampdown_length         = 0.25
        self.lr_rampup_length           = 0.05
        self.noise_ramp_length          = 0.75
        self.regularize_noise_weight    = 1e5
        self.verbose                    = verbose
        self.clone_net                  = True

        self._Gs                    = None
        self._minibatch_size        = None
        self._dlatent_avg           = None
        self._dlatent_std           = None
        self._noise_vars            = None
        self._noise_init_op         = None
        self._noise_normalize_op    = None
        self._dlatents_var          = None
        self._noise_in              = None
        self._dlatents_expr         = None
        self._images_expr           = None
        self._target_images_var     = None
        self._lpips                 = None
        self._dist                  = None
        self._loss                  = None
        self._reg_sizes             = None
        self._lrate_in              = None
        self._opt                   = None
        self._opt_step              = None
        self._cur_step              = None

    def _info(self, *args):
        if self.verbose:
            print('Projector:', *args)

    def set_network(self, Gs, minibatch_size=1):
        assert minibatch_size == 1
        self._Gs = Gs
        self._minibatch_size = minibatch_size
        if self._Gs is None:
            return
        if self.clone_net:
            self._Gs = self._Gs.clone()

        # Find dlatent stats.
        self._info('Finding W midpoint and stddev using %d samples...' % self.dlatent_avg_samples)
        latent_samples = np.random.RandomState(123).randn(self.dlatent_avg_samples, *self._Gs.input_shapes[0][1:])
        dlatent_samples = self._Gs.components.mapping.run(latent_samples, None) # [N, 18, 512]
        self._dlatent_avg = np.mean(dlatent_samples, axis=0, keepdims=True) # [1, 18, 512]
        self._dlatent_std = (np.sum((dlatent_samples - self._dlatent_avg) ** 2) / self.dlatent_avg_samples) ** 0.5
        self._info('std = %g' % self._dlatent_std)

        # Find noise inputs.
        self._info('Setting up noise inputs...')
        self._noise_vars = []
        noise_init_ops = []
        noise_normalize_ops = []
        while True:
            n = 'G_synthesis/noise%d' % len(self._noise_vars)
            if not n in self._Gs.vars:
                break
            v = self._Gs.vars[n]
            self._noise_vars.append(v)
            noise_init_ops.append(tf.assign(v, tf.random_normal(tf.shape(v), dtype=tf.float32)))
            noise_mean = tf.reduce_mean(v)
            noise_std = tf.reduce_mean((v - noise_mean)**2)**0.5
            noise_normalize_ops.append(tf.assign(v, (v - noise_mean) / noise_std))
            self._info(n, v)
        self._noise_init_op = tf.group(*noise_init_ops)
        self._noise_normalize_op = tf.group(*noise_normalize_ops)

        # Image output graph.
        self._info('Building image output graph...')
        self._dlatents_var = tf.Variable(tf.zeros([self._minibatch_size] + list(self._dlatent_avg.shape[1:])), name='dlatents_var')
        self._noise_in = tf.placeholder(tf.float32, [], name='noise_in')
        dlatents_noise = tf.random.normal(shape=self._dlatents_var.shape) * self._noise_in
        self._dlatents_expr = self._dlatents_var + dlatents_noise
        self._images_expr = self._Gs.components.synthesis.get_output_for(self._dlatents_expr, randomize_noise=False)

        # Downsample image to 256x256 if it's larger than that. VGG was built for 224x224 images.
        proc_images_expr = (self._images_expr + 1) * (255 / 2)
        sh = proc_images_expr.shape.as_list()
        if sh[2] > 256:
            factor = sh[2] // 256
            proc_images_expr = tf.reduce_mean(tf.reshape(proc_images_expr, [-1, sh[1], sh[2] // factor, factor, sh[2] // factor, factor]), axis=[3,5])

        # Loss graph.
        self._info('Building loss graph...')
        self._target_images_var = tf.Variable(tf.zeros(proc_images_expr.shape), name='target_images_var')
        if self._lpips is None:
            self._lpips = misc.load_pkl(self.vgg16_pkl) # vgg16_zhang_perceptual.pkl
        self._dist = self._lpips.get_output_for(proc_images_expr, self._target_images_var)
        self._loss = tf.reduce_sum(self._dist)

        # Noise regularization graph.
        self._info('Building noise regularization graph...')
        reg_loss = 0.0
        for v in self._noise_vars:
            sz = v.shape[2]
            while True:
                reg_loss += tf.reduce_mean(v * tf.roll(v, shift=1, axis=3))**2 + tf.reduce_mean(v * tf.roll(v, shift=1, axis=2))**2
                if sz <= 8:
                    break # Small enough already
                v = tf.reshape(v, [1, 1, sz//2, 2, sz//2, 2]) # Downscale
                v = tf.reduce_mean(v, axis=[3, 5])
                sz = sz // 2
        self._loss += reg_loss * self.regularize_noise_weight

        # Optimizer.
        self._info('Setting up optimizer...')
        self._lrate_in = tf.placeholder(tf.float32, [], name='lrate_in')
        self._opt = dnnlib.tflib.Optimizer(learning_rate=self._lrate_in)
        self._opt.register_gradients(self._loss, [self._dlatents_var] + self._noise_vars)
        self._opt_step = self._opt.apply_updates()

    def run(self, target_images):
        # Run to completion.
        self.start(target_images)
        while self._cur_step < self.num_steps:
            self.step()

        # Collect results.
        pres = dnnlib.EasyDict()
        pres.dlatents = self.get_dlatents()
        pres.noises = self.get_noises()
        pres.images = self.get_images()
        return pres

    def start(self, target_images):
        assert self._Gs is not None

        # Prepare target images.
        self._info('Preparing target images...')
        target_images = np.asarray(target_images, dtype='float32')
        target_images = (target_images + 1) * (255 / 2)
        sh = target_images.shape
        assert sh[0] == self._minibatch_size
        if sh[2] > self._target_images_var.shape[2]:
            factor = sh[2] // self._target_images_var.shape[2]
            target_images = np.reshape(target_images, [-1, sh[1], sh[2] // factor, factor, sh[3] // factor, factor]).mean((3, 5))

        # Initialize optimization state.
        self._info('Initializing optimization state...')
        tflib.set_vars({self._target_images_var: target_images, self._dlatents_var: np.tile(self._dlatent_avg, [self._minibatch_size, 1, 1])})
        tflib.run(self._noise_init_op)
        self._opt.reset_optimizer_state()
        self._cur_step = 0

    def step(self):
        assert self._cur_step is not None
        if self._cur_step >= self.num_steps:
            return
        if self._cur_step == 0:
            self._info('Running...')

        # Hyperparameters.
        t = self._cur_step / self.num_steps
        noise_strength = self._dlatent_std * self.initial_noise_factor * max(0.0, 1.0 - t / self.noise_ramp_length) ** 2
        lr_ramp = min(1.0, (1.0 - t) / self.lr_rampdown_length)
        lr_ramp = 0.5 - 0.5 * np.cos(lr_ramp * np.pi)
        lr_ramp = lr_ramp * min(1.0, t / self.lr_rampup_length)
        learning_rate = self.initial_learning_rate * lr_ramp

        # Train.
        feed_dict = {self._noise_in: noise_strength, self._lrate_in: learning_rate}
        _, dist_value, loss_value = tflib.run([self._opt_step, self._dist, self._loss], feed_dict)
        tflib.run(self._noise_normalize_op)

        # Print status.
        self._cur_step += 1
        if self._cur_step == self.num_steps or self._cur_step % 10 == 0:
            self._info('%-8d%-12g%-12g' % (self._cur_step, dist_value, loss_value))
        if self._cur_step == self.num_steps:
            self._info('Done.')

    def get_cur_step(self):
        return self._cur_step

    def get_dlatents(self):
        return tflib.run(self._dlatents_expr, {self._noise_in: 0})

    def get_noises(self):
        return tflib.run(self._noise_vars)

    def get_images(self):
        return tflib.run(self._images_expr, {self._noise_in: 0})

2.2.2. project_images

project_images.py

# from https://github.com/rolux/stylegan2encoder
import argparse
import os
import shutil
import numpy as np

import dnnlib
import dnnlib.tflib as tflib
import pretrained_networks
import projector #
import dataset_tool
from training import dataset
from training import misc


def project_image(proj, src_file, dst_dir, tmp_dir, video=False):
    data_dir = '%s/dataset' % tmp_dir
    if os.path.exists(data_dir):
        shutil.rmtree(data_dir)
    image_dir = '%s/images' % data_dir
    tfrecord_dir = '%s/tfrecords' % data_dir
    os.makedirs(image_dir, exist_ok=True)
    shutil.copy(src_file, image_dir + '/')
    #
    dataset_tool.create_from_images_raw(tfrecord_dir, image_dir, shuffle=0)
    dataset_obj = dataset.load_dataset(
        data_dir=data_dir, tfrecord_dir='tfrecords',
        max_label_size=0, repeat=False, shuffle_mb=0
    )

    print('Projecting image "%s"...' % os.path.basename(src_file))
    images, _labels = dataset_obj.get_minibatch_np(1)
    images = misc.adjust_dynamic_range(images, [0, 255], [-1, 1])
    proj.start(images)
    if video:
        video_dir = '%s/video' % tmp_dir
        os.makedirs(video_dir, exist_ok=True)
    while proj.get_cur_step() < proj.num_steps:
        print('\r%d / %d ... ' % (proj.get_cur_step(), proj.num_steps), end='', flush=True)
        proj.step()
        if video:
            filename = '%s/%08d.png' % (video_dir, proj.get_cur_step())
            misc.save_image_grid(proj.get_images(), filename, drange=[-1,1])
    print('\r%-30s\r' % '', end='', flush=True)

    os.makedirs(dst_dir, exist_ok=True)
    filename = os.path.join(dst_dir, os.path.basename(src_file)[:-4] + '.png')
    misc.save_image_grid(proj.get_images(), filename, drange=[-1,1])
    filename = os.path.join(dst_dir, os.path.basename(src_file)[:-4] + '.npy')
    np.save(filename, proj.get_dlatents()[0])


def render_video(src_file, dst_dir, tmp_dir, num_frames, mode, size, fps, codec, bitrate):
    import PIL.Image
    import moviepy.editor

    def render_frame(t):
        frame = np.clip(np.ceil(t * fps), 1, num_frames)
        image = PIL.Image.open('%s/video/%08d.png' % (tmp_dir, frame))
        if mode == 1:
            canvas = image
        else:
            canvas = PIL.Image.new('RGB', (2 * src_size, src_size))
            canvas.paste(src_image, (0, 0))
            canvas.paste(image, (src_size, 0))
        if size != src_size:
            canvas = canvas.resize((mode * size, size), PIL.Image.LANCZOS)
        return np.array(canvas)

    src_image = PIL.Image.open(src_file)
    src_size = src_image.size[1]
    duration = num_frames / fps
    filename = os.path.join(dst_dir, os.path.basename(src_file)[:-4] + '.mp4')
    video_clip = moviepy.editor.VideoClip(render_frame, duration=duration)
    video_clip.write_videofile(filename, fps=fps, codec=codec, bitrate=bitrate)


def main():
    parser = argparse.ArgumentParser(description='Project real-world images into StyleGAN2 latent space')
    parser.add_argument('src_dir', help='Directory with aligned images for projection')
    parser.add_argument('dst_dir', help='Output directory')
    parser.add_argument('--tmp-dir', default='.stylegan2-tmp', help='Temporary directory for tfrecords and video frames')
    parser.add_argument('--network-pkl', default='http://d36zk2xti64re0.cloudfront.net/stylegan2/networks/stylegan2-ffhq-config-f.pkl', help='StyleGAN2 network pickle filename')
    parser.add_argument('--vgg16-pkl', default='http://d36zk2xti64re0.cloudfront.net/stylegan1/networks/metrics/vgg16_zhang_perceptual.pkl', help='VGG16 network pickle filename')
    parser.add_argument('--num-steps', type=int, default=1000, help='Number of optimization steps')
    parser.add_argument('--initial-learning-rate', type=float, default=0.1, help='Initial learning rate')
    parser.add_argument('--initial-noise-factor', type=float, default=0.05, help='Initial noise factor')
    parser.add_argument('--verbose', type=bool, default=False, help='Verbose output')
    parser.add_argument('--video', type=bool, default=False, help='Render video of the optimization process')
    parser.add_argument('--video-mode', type=int, default=1, help='Video mode: 1 for optimization only, 2 for source + optimization')
    parser.add_argument('--video-size', type=int, default=1024, help='Video size (height in px)')
    parser.add_argument('--video-fps', type=int, default=25, help='Video framerate')
    parser.add_argument('--video-codec', default='libx264', help='Video codec')
    parser.add_argument('--video-bitrate', default='5M', help='Video bitrate')
    args = parser.parse_args()

    print('Loading networks from "%s"...' % args.network_pkl)
    _G, _D, Gs = pretrained_networks.load_networks(args.network_pkl)
    proj = projector.Projector(
        vgg16_pkl             = args.vgg16_pkl,
        num_steps             = args.num_steps,
        initial_learning_rate = args.initial_learning_rate,
        initial_noise_factor  = args.initial_noise_factor,
        verbose               = args.verbose
    )
    proj.set_network(Gs)

    src_files = sorted([os.path.join(args.src_dir, f) for f in os.listdir(args.src_dir) if f[0] not in '._'])
    for src_file in src_files:
        project_image(proj, src_file, args.dst_dir, args.tmp_dir, video=args.video)
        if args.video:
            render_video(
                src_file, args.dst_dir, args.tmp_dir, args.num_steps, args.video_mode,
                args.video_size, args.video_fps, args.video_codec, args.video_bitrate
            )
        shutil.rmtree(args.tmp_dir)

if __name__ == '__main__':
    main()

2.3. 人脸卡通化

import numpy as np
from PIL import Image
import matplotlib.pyplot as plt 
import dnnlib
import dnnlib.tflib as tflib
from pathlib import Path

latent_dir = Path("generated")
latents = latent_dir.glob("*.npy")
for latent_file in latents:
  latent = np.load(latent_file)
  latent = np.expand_dims(latent,axis=0)
  synthesis_kwargs = dict(output_transform=dict(func=tflib.convert_images_to_uint8, nchw_to_nhwc=False), minibatch_size=8)
  images = Gs_blended.components.synthesis.run(latent, randomize_noise=False, **synthesis_kwargs)
  Image.fromarray(images.transpose((0,2,3,1))[0], 'RGB').save(latent_file.parent / (f"{latent_file.stem}-toon.jpg"))

#
embedded = Image.open(filename="generated/example_01.png")
tooned = Image.open(filename="generated/example_01-toon.jpg")
plt.figure()
plt.subplot(121)
plt.imshow(embedded)
plt.axies('off')
plt.subplot(122)
plt.imshow(tooned)
plt.axies('off')
plt.show()

效果如图：

Last modification：January 5th, 2021 at 02:27 pm
© 允许规范转载