可能看過看過我上兩篇GAN和CGAN的朋友們都認為，mnist資料太簡單了，也不太適合拿出去show，所以我們來一個復雜一點的，這次難度比之前兩篇的難度又有所提升了，所以，請大家不要慌張，緊跟腳步，我們來開整，

3.PIX2PIX

3.1PIX2PIX的網路結構

可以看到仍然是兩個網路：生成器和判別器

不同的是：

1.生成器這次變樣子了，它變成了一個全卷積網路，且類似于一個U-NET網路的結構，我們一共進行了8次下采樣和8次上采樣，

2.判別這次的輸入是兩個（256，256，3）的資料了，我們把它這個兩個資料concat到一起了，這里也沒有使用全連接層了，我們也改成了全卷積，

3.2演算法邏輯

在了解演算法邏輯之前，我們先來看看這次要用到的資料集，來方便演算法邏輯的理解，

我們用到的資料集是：Facades，長什么樣子呢

就是這個樣子，每張圖片的左邊是房子的原圖，右邊是房子的簡圖，我們的目標就是：輸入簡圖，讓生成器生成和原圖相似的房子圖片，

下面我們就來盤一盤這個邏輯

1.首先給生成器輸入一個簡圖，其資料維度是N*256*256*3,經過生成器后生成一個N*256*256*3的一個資料，我們把這個生成的資料和原來的簡圖資料concat到一起，然后送入判別器，此時我們希望判別器輸出的N*16*16*1的資料是一個全1的資料，也就是說，希望生成器生成的圖能夠騙過判別器，從而來優化生成器，同時我們再讓生成的N*256*256*3的資料和它的原圖資料求一個L1loss，從而使得生成的圖和原圖又一定的相似性，這個也就是傳說中的風格遷移，

2.現在再把原圖和原始的簡圖concat到一起去，送入判別器，此時我們希望判別器的輸出N*16*16*1的資料是仍然是全一的，同時生成的圖片和原始的簡圖concat到一起去之后，希望判別器輸出的N*16*16*1的資料是全零的，以此來優化判別器，

3.至此，生成器和判別器就開始對抗起來了，印證了那句老話，我不僅要學習你，我還要超越你，

3.3實作代碼

首先是 model.py，里面定義了我們的模型結構

import torch.nn as nn
import torch


def weights_init_normal(m):
    classname = m.__class__.__name__
    if classname.find("Conv") != -1:
        torch.nn.init.normal_(m.weight.data, 0.0, 0.02)
    elif classname.find("BatchNorm2d") != -1:
        torch.nn.init.normal_(m.weight.data, 1.0, 0.02)
        torch.nn.init.constant_(m.bias.data, 0.0)


##############################
#           U-NET
##############################


class UNetDown(nn.Module):
    def __init__(self, in_size, out_size, normalize=True, dropout=0.0):
        super(UNetDown, self).__init__()
        layers = [nn.Conv2d(in_size, out_size, 4, 2, 1, bias=False)]
        if normalize:
            layers.append(nn.InstanceNorm2d(out_size))
        layers.append(nn.LeakyReLU(0.2))
        if dropout:
            layers.append(nn.Dropout(dropout))
        self.model = nn.Sequential(*layers)

    def forward(self, x):
        return self.model(x)


class UNetUp(nn.Module):
    def __init__(self, in_size, out_size, dropout=0.0):
        super(UNetUp, self).__init__()
        layers = [
            nn.ConvTranspose2d(in_size, out_size, 4, 2, 1, bias=False),
            nn.InstanceNorm2d(out_size),
            nn.ReLU(inplace=True),
        ]
        if dropout:
            layers.append(nn.Dropout(dropout))

        self.model = nn.Sequential(*layers)

    def forward(self, x, skip_input):
        x = self.model(x)
        x = torch.cat((x, skip_input), 1)

        return x


class GeneratorUNet(nn.Module):
    def __init__(self, in_channels=3, out_channels=3):
        super(GeneratorUNet, self).__init__()

        self.down1 = UNetDown(in_channels, 64, normalize=False)
        self.down2 = UNetDown(64, 128)
        self.down3 = UNetDown(128, 256)
        self.down4 = UNetDown(256, 512, dropout=0.5)
        self.down5 = UNetDown(512, 512, dropout=0.5)
        self.down6 = UNetDown(512, 512, dropout=0.5)
        self.down7 = UNetDown(512, 512, dropout=0.5)
        self.down8 = UNetDown(512, 512, normalize=False, dropout=0.5)

        self.up1 = UNetUp(512, 512, dropout=0.5)
        self.up2 = UNetUp(1024, 512, dropout=0.5)
        self.up3 = UNetUp(1024, 512, dropout=0.5)
        self.up4 = UNetUp(1024, 512, dropout=0.5)
        self.up5 = UNetUp(1024, 256)
        self.up6 = UNetUp(512, 128)
        self.up7 = UNetUp(256, 64)

        self.final = nn.Sequential(
            nn.Upsample(scale_factor=2),
            nn.ZeroPad2d((1, 0, 1, 0)),
            nn.Conv2d(128, out_channels, 4, padding=1),
            nn.Tanh(),
        )

    def forward(self, x):
        # U-Net generator with skip connections from encoder to decoder
        d1 = self.down1(x)
        d2 = self.down2(d1)
        d3 = self.down3(d2)
        d4 = self.down4(d3)
        d5 = self.down5(d4)
        d6 = self.down6(d5)
        d7 = self.down7(d6)
        d8 = self.down8(d7)
        u1 = self.up1(d8, d7)
        u2 = self.up2(u1, d6)
        u3 = self.up3(u2, d5)
        u4 = self.up4(u3, d4)
        u5 = self.up5(u4, d3)
        u6 = self.up6(u5, d2)
        u7 = self.up7(u6, d1)

        return self.final(u7)


##############################
#        Discriminator
##############################


class Discriminator(nn.Module):
    def __init__(self, in_channels=3):
        super(Discriminator, self).__init__()

        def discriminator_block(in_filters, out_filters, normalization=True):
            """Returns downsampling layers of each discriminator block"""
            layers = [nn.Conv2d(in_filters, out_filters, 4, stride=2, padding=1)]
            if normalization:
                layers.append(nn.InstanceNorm2d(out_filters))
            layers.append(nn.LeakyReLU(0.2, inplace=True))
            return layers

        self.model = nn.Sequential(
            *discriminator_block(in_channels * 2, 64, normalization=False),
            *discriminator_block(64, 128),
            *discriminator_block(128, 256),
            *discriminator_block(256, 512),
            nn.ZeroPad2d((1, 0, 1, 0)),
            nn.Conv2d(512, 1, 4, padding=1, bias=False)
        )

    def forward(self, img_A, img_B):
        # Concatenate image and condition image by channels to produce input
        img_input = torch.cat((img_A, img_B), 1)
        return self.model(img_input)

datasets.py，這個是用來讀取facades資料集的代碼

import glob
import os
import numpy as np

from torch.utils.data import Dataset
from PIL import Image
import torchvision.transforms as transforms


class ImageDataset(Dataset):
    def __init__(self, root, transforms_=None, mode="train"):
        if transforms_ is not None:
            self.transform = transforms.Compose(transforms_)
        else:
            self.transform = None

        self.files = sorted(glob.glob(os.path.join(root, mode) + "/*.*"))
        if mode == "train":
            self.files.extend(sorted(glob.glob(os.path.join(root, "test") + "/*.*")))

    def __getitem__(self, index):

        img = Image.open(self.files[index % len(self.files)])
        w, h = img.size
        img_A = img.crop((0, 0, w / 2, h))
        img_B = img.crop((w / 2, 0, w, h))

        if np.random.random() < 0.5:
            img_A = Image.fromarray(np.array(img_A)[:, ::-1, :], "RGB")
            img_B = Image.fromarray(np.array(img_B)[:, ::-1, :], "RGB")

        if self.transform is not None:
            img_A = self.transform(img_A)
            img_B = self.transform(img_B)

        return {"A": img_A, "B": img_B}

    def __len__(self):
        return len(self.files)

pix2pix.py，就是我們的訓練帶代碼了，

import argparse
import os
import numpy as np
import time
import datetime
import sys
import torch

import torchvision.transforms as transforms
from torchvision.utils import save_image

from torch.utils.data import DataLoader
from torch.autograd import Variable

from models import *
from datasets import *


parser = argparse.ArgumentParser()
parser.add_argument("--epoch", type=int, default=0, help="epoch to start training from")
parser.add_argument("--n_epochs", type=int, default=200, help="number of epochs of training")
parser.add_argument("--dataset_name", type=str, default="facades", help="name of the dataset")
parser.add_argument("--batch_size", type=int, default=1, help="size of the batches")
parser.add_argument("--lr", type=float, default=0.0002, help="adam: learning rate")
parser.add_argument("--b1", type=float, default=0.5, help="adam: decay of first order momentum of gradient")
parser.add_argument("--b2", type=float, default=0.999, help="adam: decay of first order momentum of gradient")
parser.add_argument("--decay_epoch", type=int, default=100, help="epoch from which to start lr decay")
parser.add_argument("--n_cpu", type=int, default=1, help="number of cpu threads to use during batch generation")
parser.add_argument("--img_height", type=int, default=256, help="size of image height") # 256
parser.add_argument("--img_width", type=int, default=256, help="size of image width") # 256
parser.add_argument("--channels", type=int, default=3, help="number of image channels")
parser.add_argument(
    "--sample_interval", type=int, default=500, help="interval between sampling of images from generators"
)
parser.add_argument("--checkpoint_interval", type=int, default=-1, help="interval between model checkpoints")
opt = parser.parse_args()
print(opt)

os.makedirs("images/%s" % opt.dataset_name, exist_ok=True)
os.makedirs("saved_models/%s" % opt.dataset_name, exist_ok=True)

cuda = True if torch.cuda.is_available() else False

# Loss functions
criterion_GAN = torch.nn.MSELoss()
criterion_pixelwise = torch.nn.L1Loss()

# Loss weight of L1 pixel-wise loss between translated image and real image
lambda_pixel = 100

# Calculate output of image discriminator (PatchGAN)
patch = (1, opt.img_height // 2 ** 4, opt.img_width // 2 ** 4)

# Initialize generator and discriminator
generator = GeneratorUNet()
discriminator = Discriminator()

if cuda:
    generator = generator.cuda()
    discriminator = discriminator.cuda()
    criterion_GAN.cuda()
    criterion_pixelwise.cuda()

if opt.epoch != 0:
    # Load pretrained models
    generator.load_state_dict(torch.load("saved_models/%s/generator_%d.pth" % (opt.dataset_name, opt.epoch)))
    discriminator.load_state_dict(torch.load("saved_models/%s/discriminator_%d.pth" % (opt.dataset_name, opt.epoch)))
else:
    # Initialize weights
    generator.apply(weights_init_normal)
    discriminator.apply(weights_init_normal)

# Optimizers
optimizer_G = torch.optim.Adam(generator.parameters(), lr=opt.lr, betas=(opt.b1, opt.b2))
optimizer_D = torch.optim.Adam(discriminator.parameters(), lr=opt.lr, betas=(opt.b1, opt.b2))

# Configure dataloaders
transforms_ = [
    transforms.Resize((opt.img_height, opt.img_width), Image.BICUBIC),
    transforms.ToTensor(),
    transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)),
]
import matplotlib.pyplot as plt

dataloader = DataLoader(
    ImageDataset("./data/%s" % opt.dataset_name, transforms_=transforms_),
    batch_size=opt.batch_size,
    shuffle=True,
    num_workers=opt.n_cpu,
)
val_dataloader = DataLoader(
    ImageDataset("./data/%s" % opt.dataset_name, transforms_=transforms_, mode="val"),
    batch_size=10,
    shuffle=True,
    num_workers=1,
)

# Tensor type
Tensor = torch.cuda.FloatTensor if cuda else torch.FloatTensor


def sample_images(batches_done):
    """Saves a generated sample from the validation set"""
    imgs = next(iter(val_dataloader))
    real_A = Variable(imgs["B"].type(Tensor))
    real_B = Variable(imgs["A"].type(Tensor))
    fake_B = generator(real_A)
    img_sample = torch.cat((real_A.data, fake_B.data, real_B.data), -2)
    save_image(img_sample, "images/%s/%s.png" % (opt.dataset_name, batches_done), nrow=5, normalize=True)


# ----------
#  Training
# ----------

if __name__ == '__main__':
    prev_time = time.time()

    for epoch in range(opt.epoch, opt.n_epochs):
        for i, batch in enumerate(dataloader):

            # Model inputs
            real_A = Variable(batch["B"].type(Tensor)) # batch["B"]是語意分割圖
            real_B = Variable(batch["A"].type(Tensor)) # batch["A"]是真實影像

            # Adversarial ground truths
            valid = Variable(Tensor(np.ones((real_A.size(0), *patch))), requires_grad=False)
            fake = Variable(Tensor(np.zeros((real_A.size(0), *patch))), requires_grad=False)

            # ------------------
            #  Train Generators
            # ------------------

            optimizer_G.zero_grad()

            # GAN loss
            fake_B = generator(real_A)
            pred_fake = discriminator(fake_B, real_A)
            loss_GAN = criterion_GAN(pred_fake, valid)
            # Pixel-wise loss
            loss_pixel = criterion_pixelwise(fake_B, real_B)

            # Total loss
            loss_G = loss_GAN + lambda_pixel * loss_pixel

            loss_G.backward()

            optimizer_G.step()

            # ---------------------
            #  Train Discriminator
            # ---------------------

            optimizer_D.zero_grad()

            # Real loss
            pred_real = discriminator(real_B, real_A)
            loss_real = criterion_GAN(pred_real, valid)

            # Fake loss
            pred_fake = discriminator(fake_B.detach(), real_A)
            loss_fake = criterion_GAN(pred_fake, fake)

            # Total loss
            loss_D = 0.5 * (loss_real + loss_fake)

            loss_D.backward()
            optimizer_D.step()

            # --------------
            #  Log Progress
            # --------------

            # Determine approximate time left
            batches_done = epoch * len(dataloader) + i
            batches_left = opt.n_epochs * len(dataloader) - batches_done
            time_left = datetime.timedelta(seconds=batches_left * (time.time() - prev_time))
            prev_time = time.time()

            # Print log
            sys.stdout.write(
                "\r[Epoch %d/%d] [Batch %d/%d] [D loss: %f] [G loss: %f, pixel: %f, adv: %f] ETA: %s"
                % (
                    epoch,
                    opt.n_epochs,
                    i,
                    len(dataloader),
                    loss_D.item(),
                    loss_G.item(),
                    loss_pixel.item(),
                    loss_GAN.item(),
                    time_left,
                )
            )

            # If at sample interval save image
            if batches_done % opt.sample_interval == 0:
                sample_images(batches_done)

        if opt.checkpoint_interval != -1 and epoch % opt.checkpoint_interval == 0:
            # Save model checkpoints
            torch.save(generator.state_dict(), "saved_models/%s/generator_%d.pth" % (opt.dataset_name, epoch))
            torch.save(discriminator.state_dict(), "saved_models/%s/discriminator_%d.pth" % (opt.dataset_name, epoch))

3.4效果展示

第一行，第四行是簡圖，第三行和第六行是原圖，第二行和第五行是生成圖，

3.5資料集和專案代碼

鏈接：https://pan.baidu.com/s/1xrdUO5s--5qYQ8MRbSMPRQ
提取碼：gudo

至此，敬禮，salute！！！！

老規矩，上咩咩圖