add files

gen_video.py

@@ -0,0 +1,210 @@
+import argparse
+import os
+
+import cv2
+import numpy as np
+import torch
+
+from model import Generator
+from psp_encoder.psp_encoders import PSPEncoder
+from utils import ten2cv, cv2ten
+
+import glob
+from tqdm import tqdm
+import random
+
+
+seed = 0
+
+random.seed(seed)
+np.random.seed(seed)
+torch.manual_seed(seed)
+torch.cuda.manual_seed_all(seed)
+
+
+def sigmoid(x, w=1):
+    return 1. / (1 + np.exp(-w * x))
+
+
+def get_alphas(start=-5, end=5, step=0.5, len_tail=10):
+    return [0] + [sigmoid(alpha) for alpha in np.arange(start, end, step)] + [1] * len_tail
+
+
+def slide(entries, margin=32):
+    """Returns a sliding reference window.
+    Args:
+        entries: a list containing two reference images, x_prev and x_next,
+                 both of which has a shape (1, 3, H, W)
+    Returns:
+        canvas: output slide of shape (num_frames, 3, H*2, W+margin)
+    """
+    _, C, H, W = entries[0].shape
+    alphas = get_alphas()
+    T = len(alphas) # number of frames
+
+    canvas = - torch.ones((T, C, H*2, W + margin))
+    merged = torch.cat(entries, dim=2)  # (1, 3, H*2, W)
+    for t, alpha in enumerate(alphas):
+        top = int(H * (1 - alpha))  # top, bottom for canvas
+        bottom = H * 2
+        m_top = 0  # top, bottom for merged
+        m_bottom = 2 * H - top
+        canvas[t, :, top:bottom, :W] = merged[:, :, m_top:m_bottom, :]
+    return canvas
+
+
+def slide_one_window(entries, margin=32):
+    """Returns a sliding reference window.
+    Args:
+        entries: a list containing two reference images, x_prev and x_next,
+                 both of which has a shape (1, 3, H, W)
+    Returns:
+        canvas: output slide of shape (num_frames, 3, H, W+margin)
+    """
+    _, C, H, W = entries[0].shape
+    device = entries[0].device
+    alphas = get_alphas()
+    T = len(alphas) # number of frames
+
+    canvas = - torch.ones((T, C, H, W + margin)).to(device)
+    merged = torch.cat(entries, dim=2)  # (1, 3, H*2, W)
+    for t, alpha in enumerate(alphas):
+        m_top = int(H * alpha)  # top, bottom for merged
+        m_bottom = m_top + H
+        canvas[t, :, :, :W] = merged[:, :, m_top:m_bottom, :]
+    return canvas
+
+
+def tensor2ndarray255(images):
+    images = torch.clamp(images * 0.5 + 0.5, 0, 1)
+    return (images.cpu().numpy().transpose(0, 2, 3, 1) * 255).astype(np.uint8)
+
+
+@torch.no_grad()
+def interpolate(args, g, sample_in, sample_style_prev, sample_style_next):
+    ''' returns T x C x H x W '''
+    frames_ten = []
+    alphas = get_alphas()
+
+    for alpha in alphas:
+        sample_style = torch.lerp(sample_style_prev, sample_style_next, alpha)
+        frame_ten, _ = g([sample_in], z_embed=sample_style, add_weight_index=args.add_weight_index,
+                               input_is_latent=True, return_latents=False, randomize_noise=False)
+        frames_ten.append(frame_ten)
+    frames_ten = torch.cat(frames_ten)
+    return frames_ten
+
+
+@torch.no_grad()
+def video_ref(args, g, psp_encoder, img_in_ten, img_style_tens):
+    video = []
+    sample_in = psp_encoder(img_in_ten)
+
+    img_style_ten_prev, sample_style_prev = None, None
+
+    for idx in tqdm(range(len(img_style_tens))):
+        img_style_ten_next = img_style_tens[idx]
+        sample_style_next = g_ema.get_z_embed(img_style_ten_next)
+        if img_style_ten_prev is None:
+            img_style_ten_prev, sample_style_prev = img_style_ten_next, sample_style_next
+            continue
+
+        interpolated = interpolate(args, g, sample_in, sample_style_prev, sample_style_next)
+        entries = [img_style_ten_prev, img_style_ten_next]
+        slided = slide_one_window(entries, margin=0)     # [T, C, H, W)
+        frames = torch.cat([img_in_ten.expand_as(interpolated), slided, interpolated], dim=3).cpu()   # [T, C, H, W*3)
+        video.append(frames)
+        img_style_ten_prev, sample_style_prev = img_style_ten_next, sample_style_next
+
+    # append last frame 10 time
+    for _ in range(10):
+        video.append(frames[-1:])
+    video = tensor2ndarray255(torch.cat(video))   # [T, H, W*3, C)
+
+    return video
+
+
+def save_video(fname, images, output_fps=30):
+    print('save video to: %s' % fname)
+
+    assert isinstance(images, np.ndarray), "images should be np.array: NHWC"
+    num_frames, height, width, channels = images.shape
+
+    fourcc = cv2.VideoWriter_fourcc(*'XVID')
+    videoWriter = cv2.VideoWriter(fname, fourcc, output_fps, (width, height))
+
+    for idx in tqdm(range(num_frames)):
+        frame = images[idx][:, :, ::-1]  # [H, W*3, C)
+        videoWriter.write(frame)
+
+    videoWriter.release()
+
+
+if __name__ == '__main__':
+    device = 'cuda'
+
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument('--size', type=int, default=1024)
+
+    parser.add_argument('--ckpt', type=str, default='', help='path to BlendGAN checkpoint')
+    parser.add_argument('--psp_encoder_ckpt', type=str, default='', help='path to psp_encoder checkpoint')
+
+    parser.add_argument('--style_img_path', type=str, default=None, help='path to style image')
+    parser.add_argument('--input_img_path', type=str, default=None, help='path to input image')
+    parser.add_argument('--add_weight_index', type=int, default=7)
+
+    parser.add_argument('--channel_multiplier', type=int, default=2)
+    parser.add_argument('--outdir', type=str, default="")
+
+    args = parser.parse_args()
+
+    outdir = args.outdir
+    if not os.path.exists(outdir):
+        os.makedirs(outdir, exist_ok=True)
+
+    args.latent = 512
+    args.n_mlp = 8
+
+    checkpoint = torch.load(args.ckpt)
+    model_dict = checkpoint['g_ema']
+    print('ckpt: ', args.ckpt)
+
+    g_ema = Generator(
+        args.size, args.latent, args.n_mlp, channel_multiplier=args.channel_multiplier
+    ).to(device)
+    g_ema.load_state_dict(model_dict)
+    g_ema.eval()
+
+    psp_encoder = PSPEncoder(args.psp_encoder_ckpt, output_size=args.size).to(device)
+    psp_encoder.eval()
+
+    input_img_paths = sorted(glob.glob(os.path.join(args.input_img_path, '*.*')))
+    style_img_paths = sorted(glob.glob(os.path.join(args.style_img_path, '*.*')))[:]
+
+    for input_img_path in input_img_paths:
+        print('process: %s' % input_img_path)
+
+        name_in = os.path.splitext(os.path.basename(input_img_path))[0]
+        img_in = cv2.imread(input_img_path, 1)
+        img_in = cv2.resize(img_in, (args.size, args.size))
+        img_in_ten = cv2ten(img_in, device)
+
+        img_style_tens = []
+
+        style_img_path_rand = random.choices(style_img_paths, k=8)
+        for style_img_path in style_img_path_rand:
+            name_style = os.path.splitext(os.path.basename(style_img_path))[0]
+            img_style = cv2.imread(style_img_path, 1)
+            img_style = cv2.resize(img_style, (args.size, args.size))
+            img_style_ten = cv2ten(img_style, device)
+
+            img_style_tens.append(img_style_ten)
+
+        fname = f'{args.outdir}/{name_in}.mp4'
+        video = video_ref(args, g_ema, psp_encoder, img_in_ten, img_style_tens)
+
+        save_video(fname, video, output_fps=30)
+
+    print('Done!')
+

generate_image_pairs.py

@@ -0,0 +1,111 @@
+import argparse
+import os
+
+import cv2
+import numpy as np
+import torch
+from tqdm import tqdm
+
+from model import Generator
+from utils import ten2cv, cv2ten
+import random
+
+seed = 0
+
+random.seed(seed)
+np.random.seed(seed)
+torch.manual_seed(seed)
+torch.cuda.manual_seed_all(seed)
+
+
+def generate(args, g_ema, device, mean_latent, sample_style, add_weight_index):
+    if args.sample_zs is not None:
+        sample_zs = torch.load(args.sample_zs)
+    else:
+        sample_zs = None
+
+    with torch.no_grad():
+        g_ema.eval()
+        for i in tqdm(range(args.pics)):
+            if sample_zs is not None:
+                sample_z = sample_zs[i]
+            else:
+                sample_z = torch.randn(1, args.latent, device=device)
+
+            sample1, _ = g_ema([sample_z],
+                               truncation=args.truncation, truncation_latent=mean_latent, return_latents=False, randomize_noise=False)
+            sample2, _ = g_ema([sample_z], z_embed=sample_style, add_weight_index=add_weight_index,
+                               truncation=args.truncation, truncation_latent=mean_latent, return_latents=False, randomize_noise=False)
+
+            sample1 = ten2cv(sample1)
+            sample2 = ten2cv(sample2)
+            out = np.concatenate([sample1, sample2], axis=1)
+
+            cv2.imwrite(f'{args.outdir}/{str(i).zfill(6)}.jpg', out)
+
+
+if __name__ == '__main__':
+    device = 'cuda'
+
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument('--size', type=int, default=1024)
+    parser.add_argument('--pics', type=int, default=20, help='N_PICS')
+    parser.add_argument('--truncation', type=float, default=0.75)
+    parser.add_argument('--truncation_mean', type=int, default=4096)
+    parser.add_argument('--ckpt', type=str, default='', help='path to BlendGAN checkpoint')
+    parser.add_argument('--style_img', type=str, default=None, help='path to style image')
+    parser.add_argument('--sample_zs', type=str, default=None)
+    parser.add_argument('--add_weight_index', type=int, default=6)
+
+    parser.add_argument('--channel_multiplier', type=int, default=2)
+    parser.add_argument('--outdir', type=str, default="")
+
+    args = parser.parse_args()
+
+    outdir = args.outdir
+    if not os.path.exists(outdir):
+        os.makedirs(outdir, exist_ok=True)
+
+    args.latent = 512
+    args.n_mlp = 8
+
+    checkpoint = torch.load(args.ckpt)
+    model_dict = checkpoint['g_ema']
+    if "latent_avg" in checkpoint.keys():
+        latent_avg = checkpoint["latent_avg"]
+    else:
+        latent_avg = None
+    if "truncation" in checkpoint.keys():
+        args.truncation = checkpoint["truncation"]
+
+    print('ckpt: ', args.ckpt)
+    print('truncation: ', args.truncation)
+
+    g_ema = Generator(
+        args.size, args.latent, args.n_mlp, channel_multiplier=args.channel_multiplier
+    ).to(device)
+    g_ema.load_state_dict(model_dict)
+
+    if args.truncation < 1:
+        if latent_avg is not None:
+            mean_latent = latent_avg
+            print('### use mean_latent in ckpt["latent_avg"]')
+        else:
+            with torch.no_grad():
+                mean_latent = g_ema.mean_latent(args.truncation_mean)
+                print('### generate mean_latent with \'g_ema.mean_latent\'')
+    else:
+        mean_latent = None
+        print('### args.truncation = 1, mean_latent is None')
+
+    if args.style_img is not None:
+        img = cv2.imread(args.style_img, 1)
+        img = cv2ten(img, device)
+        sample_style = g_ema.get_z_embed(img)
+    else:
+        sample_style = torch.randn(1, args.latent, device=device)
+
+    generate(args, g_ema, device, mean_latent, sample_style, args.add_weight_index)
+
+    print('Done!')

model.py

@@ -0,0 +1,782 @@
+import math
+import random
+
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+from op import FusedLeakyReLU, fused_leaky_relu, upfirdn2d
+from model_encoder import StyleEncoder
+
+
+class PixelNorm(nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, input):
+        return input * torch.rsqrt(torch.mean(input ** 2, dim=1, keepdim=True) + 1e-8)
+
+
+def make_kernel(k):
+    k = torch.tensor(k, dtype=torch.float32)
+
+    if k.ndim == 1:
+        k = k[None, :] * k[:, None]
+
+    k /= k.sum()
+
+    return k
+
+
+class Upsample(nn.Module):
+    def __init__(self, kernel, factor=2):
+        super().__init__()
+
+        self.factor = factor
+        kernel = make_kernel(kernel) * (factor ** 2)
+        self.register_buffer('kernel', kernel)
+
+        p = kernel.shape[0] - factor
+
+        pad0 = (p + 1) // 2 + factor - 1
+        pad1 = p // 2
+
+        self.pad = (pad0, pad1)
+
+    def forward(self, input):
+        out = upfirdn2d(input, self.kernel, up=self.factor, down=1, pad=self.pad)
+
+        return out
+
+
+class Downsample(nn.Module):
+    def __init__(self, kernel, factor=2):
+        super().__init__()
+
+        self.factor = factor
+        kernel = make_kernel(kernel)
+        self.register_buffer('kernel', kernel)
+
+        p = kernel.shape[0] - factor
+
+        pad0 = (p + 1) // 2
+        pad1 = p // 2
+
+        self.pad = (pad0, pad1)
+
+    def forward(self, input):
+        out = upfirdn2d(input, self.kernel, up=1, down=self.factor, pad=self.pad)
+
+        return out
+
+
+class Blur(nn.Module):
+    def __init__(self, kernel, pad, upsample_factor=1):
+        super().__init__()
+
+        kernel = make_kernel(kernel)
+
+        if upsample_factor > 1:
+            kernel = kernel * (upsample_factor ** 2)
+
+        self.register_buffer('kernel', kernel)
+
+        self.pad = pad
+
+    def forward(self, input):
+        out = upfirdn2d(input, self.kernel, pad=self.pad)
+
+        return out
+
+
+class EqualConv2d(nn.Module):
+    def __init__(
+        self, in_channel, out_channel, kernel_size, stride=1, padding=0, bias=True
+    ):
+        super().__init__()
+
+        self.weight = nn.Parameter(
+            torch.randn(out_channel, in_channel, kernel_size, kernel_size)
+        )
+        self.scale = 1 / math.sqrt(in_channel * kernel_size ** 2)
+
+        self.stride = stride
+        self.padding = padding
+
+        if bias:
+            self.bias = nn.Parameter(torch.zeros(out_channel))
+
+        else:
+            self.bias = None
+
+    def forward(self, input):
+        out = F.conv2d(
+            input,
+            self.weight * self.scale,
+            bias=self.bias,
+            stride=self.stride,
+            padding=self.padding,
+        )
+
+        return out
+
+    def __repr__(self):
+        return (
+            f'{self.__class__.__name__}({self.weight.shape[1]}, {self.weight.shape[0]},'
+            f' {self.weight.shape[2]}, stride={self.stride}, padding={self.padding})'
+        )
+
+
+class EqualLinear(nn.Module):
+    def __init__(
+        self, in_dim, out_dim, bias=True, bias_init=0, lr_mul=1, activation=None
+    ):
+        super().__init__()
+
+        self.weight = nn.Parameter(torch.randn(out_dim, in_dim).div_(lr_mul))
+
+        if bias:
+            self.bias = nn.Parameter(torch.zeros(out_dim).fill_(bias_init))
+
+        else:
+            self.bias = None
+
+        self.activation = activation
+
+        self.scale = (1 / math.sqrt(in_dim)) * lr_mul
+        self.lr_mul = lr_mul
+
+    def forward(self, input):
+        if self.activation:
+            out = F.linear(input, self.weight * self.scale)
+            out = fused_leaky_relu(out, self.bias * self.lr_mul)
+
+        else:
+            out = F.linear(
+                input, self.weight * self.scale, bias=self.bias * self.lr_mul
+            )
+
+        return out
+
+    def __repr__(self):
+        return (
+            f'{self.__class__.__name__}({self.weight.shape[1]}, {self.weight.shape[0]})'
+        )
+
+
+class ScaledLeakyReLU(nn.Module):
+    def __init__(self, negative_slope=0.2):
+        super().__init__()
+
+        self.negative_slope = negative_slope
+
+    def forward(self, input):
+        out = F.leaky_relu(input, negative_slope=self.negative_slope)
+
+        return out * math.sqrt(2)
+
+
+class ModulatedConv2d(nn.Module):
+    def __init__(
+        self,
+        in_channel,
+        out_channel,
+        kernel_size,
+        style_dim,
+        demodulate=True,
+        upsample=False,
+        downsample=False,
+        blur_kernel=[1, 3, 3, 1],
+    ):
+        super().__init__()
+
+        self.eps = 1e-8
+        self.kernel_size = kernel_size
+        self.in_channel = in_channel
+        self.out_channel = out_channel
+        self.upsample = upsample
+        self.downsample = downsample
+
+        if upsample:
+            factor = 2
+            p = (len(blur_kernel) - factor) - (kernel_size - 1)
+            pad0 = (p + 1) // 2 + factor - 1
+            pad1 = p // 2 + 1
+
+            self.blur = Blur(blur_kernel, pad=(pad0, pad1), upsample_factor=factor)
+
+        if downsample:
+            factor = 2
+            p = (len(blur_kernel) - factor) + (kernel_size - 1)
+            pad0 = (p + 1) // 2
+            pad1 = p // 2
+
+            self.blur = Blur(blur_kernel, pad=(pad0, pad1))
+
+        fan_in = in_channel * kernel_size ** 2
+        self.scale = 1 / math.sqrt(fan_in)
+        self.padding = kernel_size // 2
+
+        self.weight = nn.Parameter(
+            torch.randn(1, out_channel, in_channel, kernel_size, kernel_size)
+        )
+
+        self.modulation = EqualLinear(style_dim, in_channel, bias_init=1)
+
+        self.demodulate = demodulate
+
+    def __repr__(self):
+        return (
+            f'{self.__class__.__name__}({self.in_channel}, {self.out_channel}, {self.kernel_size}, '
+            f'upsample={self.upsample}, downsample={self.downsample})'
+        )
+
+    def forward(self, input, style):
+        batch, in_channel, height, width = input.shape
+
+        style = self.modulation(style).view(batch, 1, in_channel, 1, 1)
+        weight = self.scale * self.weight * style
+
+        if self.demodulate:
+            demod = torch.rsqrt(weight.pow(2).sum([2, 3, 4]) + 1e-8)
+            weight = weight * demod.view(batch, self.out_channel, 1, 1, 1)
+
+        weight = weight.view(
+            batch * self.out_channel, in_channel, self.kernel_size, self.kernel_size
+        )
+
+        if self.upsample:
+            input = input.view(1, batch * in_channel, height, width)
+            weight = weight.view(
+                batch, self.out_channel, in_channel, self.kernel_size, self.kernel_size
+            )
+            weight = weight.transpose(1, 2).reshape(
+                batch * in_channel, self.out_channel, self.kernel_size, self.kernel_size
+            )
+            out = F.conv_transpose2d(input, weight, padding=0, stride=2, groups=batch)
+            _, _, height, width = out.shape
+            out = out.view(batch, self.out_channel, height, width)
+            out = self.blur(out)
+
+        elif self.downsample:
+            input = self.blur(input)
+            _, _, height, width = input.shape
+            input = input.view(1, batch * in_channel, height, width)
+            out = F.conv2d(input, weight, padding=0, stride=2, groups=batch)
+            _, _, height, width = out.shape
+            out = out.view(batch, self.out_channel, height, width)
+
+        else:
+            input = input.view(1, batch * in_channel, height, width)
+            out = F.conv2d(input, weight, padding=self.padding, groups=batch)
+            _, _, height, width = out.shape
+            out = out.view(batch, self.out_channel, height, width)
+
+        return out
+
+
+class NoiseInjection(nn.Module):
+    def __init__(self):
+        super().__init__()
+
+        self.weight = nn.Parameter(torch.zeros(1))
+
+    def forward(self, image, noise=None):
+        if noise is None:
+            batch, _, height, width = image.shape
+            noise = image.new_empty(batch, 1, height, width).normal_()
+
+        return image + self.weight * noise
+
+
+class ConstantInput(nn.Module):
+    def __init__(self, channel, size=4):
+        super().__init__()
+
+        self.input = nn.Parameter(torch.randn(1, channel, size, size))
+
+    def forward(self, input):
+        batch = input.shape[0]
+        out = self.input.repeat(batch, 1, 1, 1)
+
+        return out
+
+
+class StyledConv(nn.Module):
+    def __init__(
+        self,
+        in_channel,
+        out_channel,
+        kernel_size,
+        style_dim,
+        upsample=False,
+        blur_kernel=[1, 3, 3, 1],
+        demodulate=True,
+    ):
+        super().__init__()
+
+        self.conv = ModulatedConv2d(
+            in_channel,
+            out_channel,
+            kernel_size,
+            style_dim,
+            upsample=upsample,
+            blur_kernel=blur_kernel,
+            demodulate=demodulate,
+        )
+
+        self.noise = NoiseInjection()
+        # self.bias = nn.Parameter(torch.zeros(1, out_channel, 1, 1))
+        # self.activate = ScaledLeakyReLU(0.2)
+        self.activate = FusedLeakyReLU(out_channel)
+
+    def forward(self, input, style, noise=None):
+        out = self.conv(input, style)
+        out = self.noise(out, noise=noise)
+        # out = out + self.bias
+        out = self.activate(out)
+
+        return out
+
+
+class ToRGB(nn.Module):
+    def __init__(self, in_channel, style_dim, upsample=True, blur_kernel=[1, 3, 3, 1]):
+        super().__init__()
+
+        if upsample:
+            self.upsample = Upsample(blur_kernel)
+
+        self.conv = ModulatedConv2d(in_channel, 3, 1, style_dim, demodulate=False)
+        self.bias = nn.Parameter(torch.zeros(1, 3, 1, 1))
+
+    def forward(self, input, style, skip=None):
+        out = self.conv(input, style)
+        out = out + self.bias
+
+        if skip is not None:
+            skip = self.upsample(skip)
+
+            out = out + skip
+
+        return out
+
+
+class Generator(nn.Module):
+    def __init__(
+        self,
+        size,
+        style_dim,
+        n_mlp,
+        channel_multiplier=2,
+        blur_kernel=[1, 3, 3, 1],
+        lr_mlp=0.01,
+    ):
+        super().__init__()
+
+        self.size = size
+
+        self.style_dim = style_dim
+
+        self.embedder = StyleEncoder(style_dim=512, n_mlp=4)
+
+        layers = [PixelNorm()]
+
+        for i in range(n_mlp):
+            layers.append(
+                EqualLinear(
+                    style_dim, style_dim, lr_mul=lr_mlp, activation='fused_lrelu'
+                )
+            )
+        self.embedding = nn.Sequential(*layers)
+
+        layers = [PixelNorm()]
+
+        for i in range(n_mlp):
+            layers.append(
+                EqualLinear(
+                    style_dim, style_dim, lr_mul=lr_mlp, activation='fused_lrelu'
+                )
+            )
+
+        self.style = nn.Sequential(*layers)
+
+        self.channels = {
+            4: 512,
+            8: 512,
+            16: 512,
+            32: 512,
+            64: 256 * channel_multiplier,
+            128: 128 * channel_multiplier,
+            256: 64 * channel_multiplier,
+            512: 32 * channel_multiplier,
+            1024: 16 * channel_multiplier,
+        }
+
+        self.input = ConstantInput(self.channels[4])
+        self.conv1 = StyledConv(
+            self.channels[4], self.channels[4], 3, style_dim, blur_kernel=blur_kernel
+        )
+        self.to_rgb1 = ToRGB(self.channels[4], style_dim, upsample=False)
+
+        self.log_size = int(math.log(size, 2))
+        self.num_layers = (self.log_size - 2) * 2 + 1
+
+        self.convs = nn.ModuleList()
+        self.upsamples = nn.ModuleList()
+        self.to_rgbs = nn.ModuleList()
+        self.noises = nn.Module()
+
+        in_channel = self.channels[4]
+
+        for layer_idx in range(self.num_layers):
+            res = (layer_idx + 5) // 2
+            shape = [1, 1, 2 ** res, 2 ** res]
+            self.noises.register_buffer(f'noise_{layer_idx}', torch.randn(*shape))
+
+        for i in range(3, self.log_size + 1):
+            out_channel = self.channels[2 ** i]
+
+            self.convs.append(
+                StyledConv(
+                    in_channel,
+                    out_channel,
+                    3,
+                    style_dim,
+                    upsample=True,
+                    blur_kernel=blur_kernel,
+                )
+            )
+
+            self.convs.append(
+                StyledConv(
+                    out_channel, out_channel, 3, style_dim, blur_kernel=blur_kernel
+                )
+            )
+
+            self.to_rgbs.append(ToRGB(out_channel, style_dim))
+
+            in_channel = out_channel
+
+        self.n_latent = self.log_size * 2 - 2
+
+        self.add_weight = nn.Parameter(torch.ones(1, self.n_latent, 1))
+
+    def make_noise(self):
+        device = self.input.input.device
+
+        noises = [torch.randn(1, 1, 2 ** 2, 2 ** 2, device=device)]
+
+        for i in range(3, self.log_size + 1):
+            for _ in range(2):
+                noises.append(torch.randn(1, 1, 2 ** i, 2 ** i, device=device))
+
+        return noises
+
+    def mean_latent(self, n_latent):
+        latent_in = torch.randn(
+            n_latent, self.style_dim, device=self.input.input.device
+        )
+        latent = self.style(latent_in).mean(0, keepdim=True)
+
+        return latent
+
+    def get_latent(self, input):
+        return self.style(input)
+
+    def get_z_embed(self, image):
+        self.embedder.eval()
+        with torch.no_grad():
+            z_embed = self.embedder(image)  # [N, 512]
+        return z_embed
+
+    def forward(
+        self,
+        styles=None,
+        return_latents=False,
+        inject_index=None,
+        truncation=1,
+        truncation_latent=None,
+        input_is_latent=False,
+        noise=None,
+        randomize_noise=True,
+        style_image=None,
+        z_embed=None,
+        only_return_z_embed=False,
+        add_weight_index=None,
+    ):
+        if only_return_z_embed and style_image is not None:
+            return self.get_z_embed(style_image)
+
+        if not input_is_latent:
+            styles = [self.style(s) for s in styles]
+
+        if noise is None:
+            if randomize_noise:
+                noise = [None] * self.num_layers
+            else:
+                noise = [
+                    getattr(self.noises, f'noise_{i}') for i in range(self.num_layers)
+                ]
+
+        if truncation < 1:
+            style_t = []
+
+            for style in styles:
+                style_t.append(
+                    truncation_latent + truncation * (style - truncation_latent)
+                )
+
+            styles = style_t
+
+        if len(styles) < 2:
+            inject_index = self.n_latent
+
+            if styles[0].ndim < 3:
+                latent = styles[0].unsqueeze(1).repeat(1, inject_index, 1)
+
+            else:
+                latent = styles[0]
+
+        else:
+            if inject_index is None:
+                inject_index = random.randint(1, self.n_latent - 1)
+
+            latent = styles[0].unsqueeze(1).repeat(1, inject_index, 1)
+            latent2 = styles[1].unsqueeze(1).repeat(1, self.n_latent - inject_index, 1)
+
+            latent = torch.cat([latent, latent2], 1)  # [N, 18, 512]
+
+        if z_embed is not None:
+            latent_style = self.embedding(z_embed)
+            latent_style = latent_style.unsqueeze(1).repeat(1, self.n_latent, 1)  # [N, 18, 512]
+        elif style_image is not None:
+            z_embed = self.get_z_embed(style_image)  # [N, 512]
+            latent_style = self.embedding(z_embed)
+            latent_style = latent_style.unsqueeze(1).repeat(1, self.n_latent, 1)  # [N, 18, 512]
+        else:
+            latent_style = None
+
+        if latent_style is not None:
+            self.add_weight.data = self.add_weight.data.clamp(0.0, 1.0)
+            if add_weight_index is not None:
+                add_weight_new = self.add_weight.clone()
+                add_weight_new[:, :add_weight_index, :] = 1.0
+                latent = latent * add_weight_new + latent_style * (1 - add_weight_new)
+            else:
+                latent = latent * self.add_weight + latent_style * (1 - self.add_weight)
+
+        out = self.input(latent)
+        out = self.conv1(out, latent[:, 0], noise=noise[0])
+
+        skip = self.to_rgb1(out, latent[:, 1])
+
+        i = 1
+        for conv1, conv2, noise1, noise2, to_rgb in zip(
+            self.convs[::2], self.convs[1::2], noise[1::2], noise[2::2], self.to_rgbs
+        ):
+            out = conv1(out, latent[:, i], noise=noise1)
+            out = conv2(out, latent[:, i + 1], noise=noise2)
+            skip = to_rgb(out, latent[:, i + 2], skip)
+
+            i += 2
+
+        image = skip
+
+        if return_latents:
+            if style_image is not None or z_embed is not None:
+                return image, latent, z_embed
+            else:
+                return image, latent
+
+        else:
+            return image, None
+
+
+class ConvLayer(nn.Sequential):
+    def __init__(
+        self,
+        in_channel,
+        out_channel,
+        kernel_size,
+        downsample=False,
+        blur_kernel=[1, 3, 3, 1],
+        bias=True,
+        activate=True,
+    ):
+        layers = []
+
+        if downsample:
+            factor = 2
+            p = (len(blur_kernel) - factor) + (kernel_size - 1)
+            pad0 = (p + 1) // 2
+            pad1 = p // 2
+
+            layers.append(Blur(blur_kernel, pad=(pad0, pad1)))
+
+            stride = 2
+            self.padding = 0
+
+        else:
+            stride = 1
+            self.padding = kernel_size // 2
+
+        layers.append(
+            EqualConv2d(
+                in_channel,
+                out_channel,
+                kernel_size,
+                padding=self.padding,
+                stride=stride,
+                bias=bias and not activate,
+            )
+        )
+
+        if activate:
+            if bias:
+                layers.append(FusedLeakyReLU(out_channel))
+
+            else:
+                layers.append(ScaledLeakyReLU(0.2))
+
+        super().__init__(*layers)
+
+
+class ResBlock(nn.Module):
+    def __init__(self, in_channel, out_channel, blur_kernel=[1, 3, 3, 1]):
+        super().__init__()
+
+        self.conv1 = ConvLayer(in_channel, in_channel, 3)
+        self.conv2 = ConvLayer(in_channel, out_channel, 3, downsample=True)
+
+        self.skip = ConvLayer(
+            in_channel, out_channel, 1, downsample=True, activate=False, bias=False
+        )
+
+    def forward(self, input):
+        out = self.conv1(input)
+        out = self.conv2(out)
+
+        skip = self.skip(input)
+        out = (out + skip) / math.sqrt(2)
+
+        return out
+
+
+class Discriminator(nn.Module):
+    def __init__(self, size, channel_multiplier=2, blur_kernel=[1, 3, 3, 1]):
+        super().__init__()
+
+        channels = {
+            4: 512,
+            8: 512,
+            16: 512,
+            32: 512,
+            64: 256 * channel_multiplier,
+            128: 128 * channel_multiplier,
+            256: 64 * channel_multiplier,
+            512: 32 * channel_multiplier,
+            1024: 16 * channel_multiplier,
+        }
+
+        convs = [ConvLayer(3, channels[size], 1)]
+
+        log_size = int(math.log(size, 2))
+
+        in_channel = channels[size]
+
+        for i in range(log_size, 2, -1):
+            out_channel = channels[2 ** (i - 1)]
+
+            convs.append(ResBlock(in_channel, out_channel, blur_kernel))
+
+            in_channel = out_channel
+
+        self.convs = nn.Sequential(*convs)
+
+        self.stddev_group = 4
+        self.stddev_feat = 1
+
+        self.final_conv = ConvLayer(in_channel + 1, channels[4], 3)
+        self.final_linear = nn.Sequential(
+            EqualLinear(channels[4] * 4 * 4, channels[4], activation='fused_lrelu'),
+            EqualLinear(channels[4], 1),
+        )
+
+    def forward(self, input):
+        out = self.convs(input)
+
+        batch, channel, height, width = out.shape
+        group = min(batch, self.stddev_group)
+        stddev = out.view(
+            group, -1, self.stddev_feat, channel // self.stddev_feat, height, width
+        )
+        stddev = torch.sqrt(stddev.var(0, unbiased=False) + 1e-8)
+        stddev = stddev.mean([2, 3, 4], keepdims=True).squeeze(2)
+        stddev = stddev.repeat(group, 1, height, width)
+        out = torch.cat([out, stddev], 1)
+
+        out = self.final_conv(out)
+
+        out = out.view(batch, -1)
+        out = self.final_linear(out)
+
+        return out
+
+
+class ProjectionDiscriminator(nn.Module):
+    def __init__(self, size, style_dim=512, channel_multiplier=2, blur_kernel=[1, 3, 3, 1]):
+        super().__init__()
+
+        channels = {
+            4: 512,
+            8: 512,
+            16: 512,
+            32: 512,
+            64: 256 * channel_multiplier,
+            128: 128 * channel_multiplier,
+            256: 64 * channel_multiplier,
+            512: 32 * channel_multiplier,
+            1024: 16 * channel_multiplier,
+        }
+
+        convs = [ConvLayer(3, channels[size], 1)]
+
+        log_size = int(math.log(size, 2))
+
+        in_channel = channels[size]
+
+        for i in range(log_size, 2, -1):
+            out_channel = channels[2 ** (i - 1)]
+            convs.append(ResBlock(in_channel, out_channel, blur_kernel))
+            in_channel = out_channel
+
+        self.convs = nn.Sequential(*convs)
+
+        self.stddev_group = 4
+        self.stddev_feat = 1
+
+        self.final_conv = ConvLayer(in_channel + 1, channels[4], 3)
+
+        self.l_out = EqualLinear(in_channel, 1)
+        self.l_style = EqualLinear(style_dim, in_channel)
+
+    def forward(self, input, style):
+        out = self.convs(input)
+
+        batch, channel, height, width = out.shape
+        group = min(batch, self.stddev_group)
+        stddev = out.view(
+            group, -1, self.stddev_feat, channel // self.stddev_feat, height, width
+        )
+        stddev = torch.sqrt(stddev.var(0, unbiased=False) + 1e-8)
+        stddev = stddev.mean([2, 3, 4], keepdims=True).squeeze(2)
+        stddev = stddev.repeat(group, 1, height, width)
+        out = torch.cat([out, stddev], 1)
+
+        out = self.final_conv(out)
+
+        h = torch.sum(out, dim=(2, 3))
+        output = self.l_out(h)
+        output += torch.sum(self.l_style(style) * h, dim=1, keepdim=True)
+
+        return output

model_encoder.py

@@ -0,0 +1,160 @@
+import math
+
+from collections import namedtuple
+
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+import torchvision.models.vgg as vgg
+
+from op import fused_leaky_relu
+
+
+FeatureOutput = namedtuple(
+    "FeatureOutput", ["relu1", "relu2", "relu3", "relu4", "relu5"])
+
+
+def gram_matrix(y):
+    (b, ch, h, w) = y.size()
+    features = y.view(b, ch, w * h)
+    features_t = features.transpose(1, 2)
+    gram = features.bmm(features_t) / (ch * h * w)
+    return gram
+
+
+class FeatureExtractor(nn.Module):
+    """Reference:
+        https://discuss.pytorch.org/t/how-to-extract-features-of-an-image-from-a-trained-model/119/3
+    """
+
+    def __init__(self):
+        super(FeatureExtractor, self).__init__()
+        self.vgg_layers = vgg.vgg19(pretrained=True).features
+        self.layer_name_mapping = {
+            '3': "relu1",
+            '8': "relu2",
+            '17': "relu3",
+            '26': "relu4",
+            '35': "relu5",
+        }
+
+    def forward(self, x):
+        output = {}
+        for name, module in self.vgg_layers._modules.items():
+            x = module(x)
+            if name in self.layer_name_mapping:
+                output[self.layer_name_mapping[name]] = x
+        return FeatureOutput(**output)
+
+
+class StyleEmbedder(nn.Module):
+    def __init__(self):
+        super(StyleEmbedder, self).__init__()
+        self.feature_extractor = FeatureExtractor()
+        self.feature_extractor.eval()
+        self.avg_pool = torch.nn.AdaptiveAvgPool2d((256, 256))
+
+    def forward(self, img):
+        N = img.shape[0]
+        features = self.feature_extractor(self.avg_pool(img))
+
+        grams = []
+        for feature in features:
+            gram = gram_matrix(feature)
+            grams.append(gram.view(N, -1))
+        out = torch.cat(grams, dim=1)
+        return out
+
+
+class PixelNorm(nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, input):
+        return input * torch.rsqrt(torch.mean(input ** 2, dim=1, keepdim=True) + 1e-8)
+
+
+class EqualLinear(nn.Module):
+    def __init__(
+        self, in_dim, out_dim, bias=True, bias_init=0, lr_mul=1, activation=None
+    ):
+        super().__init__()
+
+        self.weight = nn.Parameter(torch.randn(out_dim, in_dim).div_(lr_mul))
+
+        if bias:
+            self.bias = nn.Parameter(torch.zeros(out_dim).fill_(bias_init))
+
+        else:
+            self.bias = None
+
+        self.activation = activation
+
+        self.scale = (1 / math.sqrt(in_dim)) * lr_mul
+        self.lr_mul = lr_mul
+
+    def forward(self, input):
+        if self.activation:
+            out = F.linear(input, self.weight * self.scale)
+            out = fused_leaky_relu(out, self.bias * self.lr_mul)
+
+        else:
+            out = F.linear(
+                input, self.weight * self.scale, bias=self.bias * self.lr_mul
+            )
+
+        return out
+
+    def __repr__(self):
+        return (
+            f'{self.__class__.__name__}({self.weight.shape[1]}, {self.weight.shape[0]})'
+        )
+
+
+class StyleEncoder(nn.Module):
+    def __init__(
+        self,
+        style_dim=512,
+        n_mlp=4,
+    ):
+        super().__init__()
+
+        self.style_dim = style_dim
+
+        e_dim = 610304
+        self.embedder = StyleEmbedder()
+
+        layers = []
+
+        layers.append(EqualLinear(e_dim, style_dim, lr_mul=1, activation='fused_lrelu'))
+        for i in range(n_mlp - 2):
+            layers.append(
+                EqualLinear(
+                    style_dim, style_dim, lr_mul=1, activation='fused_lrelu'
+                )
+            )
+        layers.append(EqualLinear(style_dim, style_dim, lr_mul=1, activation=None))
+        self.embedder_mlp = nn.Sequential(*layers)
+
+    def forward(self, image):
+        z_embed = self.embedder_mlp(self.embedder(image))  # [N, 512]
+        return z_embed
+
+
+class Projector(nn.Module):
+    def __init__(self, style_dim=512, n_mlp=4):
+        super().__init__()
+
+        layers = []
+        for i in range(n_mlp - 1):
+            layers.append(
+                EqualLinear(
+                    style_dim, style_dim, lr_mul=1, activation='fused_lrelu'
+                )
+            )
+        layers.append(EqualLinear(style_dim, style_dim, lr_mul=1, activation=None))
+        self.projector = nn.Sequential(*layers)
+
+    def forward(self, x):
+        return self.projector(x)

op/__init__.py

@@ -0,0 +1,2 @@
+from .fused_act import FusedLeakyReLU, fused_leaky_relu
+from .upfirdn2d import upfirdn2d

op/fused_act.py

@@ -0,0 +1,104 @@
+import os
+
+import torch
+from torch import nn
+from torch.nn import functional as F
+from torch.autograd import Function
+from torch.utils.cpp_extension import load
+
+
+module_path = os.path.dirname(__file__)
+
+cuda_available = torch.cuda.is_available()
+
+if cuda_available:
+    fused = load(
+        "fused",
+        sources=[
+            os.path.join(module_path, "fused_bias_act.cpp"),
+            os.path.join(module_path, "fused_bias_act_kernel.cu"),
+        ],
+    )
+else:
+    fused = None
+    print("fused_act.py is running on cpu")
+
+
+class FusedLeakyReLUFunctionBackward(Function):
+    @staticmethod
+    def forward(ctx, grad_output, out, negative_slope, scale):
+        ctx.save_for_backward(out)
+        ctx.negative_slope = negative_slope
+        ctx.scale = scale
+
+        empty = grad_output.new_empty(0)
+
+        grad_input = fused.fused_bias_act(
+            grad_output, empty, out, 3, 1, negative_slope, scale
+        )
+
+        dim = [0]
+
+        if grad_input.ndim > 2:
+            dim += list(range(2, grad_input.ndim))
+
+        grad_bias = grad_input.sum(dim).detach()
+
+        return grad_input, grad_bias
+
+    @staticmethod
+    def backward(ctx, gradgrad_input, gradgrad_bias):
+        out, = ctx.saved_tensors
+        gradgrad_out = fused.fused_bias_act(
+            gradgrad_input, gradgrad_bias, out, 3, 1, ctx.negative_slope, ctx.scale
+        )
+
+        return gradgrad_out, None, None, None
+
+
+class FusedLeakyReLUFunction(Function):
+    @staticmethod
+    def forward(ctx, input, bias, negative_slope, scale):
+        empty = input.new_empty(0)
+        out = fused.fused_bias_act(input, bias, empty, 3, 0, negative_slope, scale)
+        ctx.save_for_backward(out)
+        ctx.negative_slope = negative_slope
+        ctx.scale = scale
+
+        return out
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        out, = ctx.saved_tensors
+
+        grad_input, grad_bias = FusedLeakyReLUFunctionBackward.apply(
+            grad_output, out, ctx.negative_slope, ctx.scale
+        )
+
+        return grad_input, grad_bias, None, None
+
+
+class FusedLeakyReLU(nn.Module):
+    def __init__(self, channel, negative_slope=0.2, scale=2 ** 0.5):
+        super().__init__()
+
+        self.bias = nn.Parameter(torch.zeros(channel))
+        self.negative_slope = negative_slope
+        self.scale = scale
+
+    def forward(self, input):
+        return fused_leaky_relu(input, self.bias, self.negative_slope, self.scale)
+
+
+def fused_leaky_relu(input, bias, negative_slope=0.2, scale=2 ** 0.5):
+    if input.device.type == "cpu":
+        rest_dim = [1] * (input.ndim - bias.ndim - 1)
+        return (
+            F.leaky_relu(
+                input + bias.view(1, bias.shape[0], *rest_dim), negative_slope=0.2
+            )
+            * scale
+        )
+
+    else:
+        return FusedLeakyReLUFunction.apply(input, bias, negative_slope, scale)

op/fused_bias_act.cpp

@@ -0,0 +1,21 @@
+#include <torch/extension.h>
+
+
+torch::Tensor fused_bias_act_op(const torch::Tensor& input, const torch::Tensor& bias, const torch::Tensor& refer,
+    int act, int grad, float alpha, float scale);
+
+#define CHECK_CUDA(x) TORCH_CHECK(x.type().is_cuda(), #x " must be a CUDA tensor")
+#define CHECK_CONTIGUOUS(x) TORCH_CHECK(x.is_contiguous(), #x " must be contiguous")
+#define CHECK_INPUT(x) CHECK_CUDA(x); CHECK_CONTIGUOUS(x)
+
+torch::Tensor fused_bias_act(const torch::Tensor& input, const torch::Tensor& bias, const torch::Tensor& refer,
+    int act, int grad, float alpha, float scale) {
+    CHECK_CUDA(input);
+    CHECK_CUDA(bias);
+
+    return fused_bias_act_op(input, bias, refer, act, grad, alpha, scale);
+}
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+    m.def("fused_bias_act", &fused_bias_act, "fused bias act (CUDA)");
+}