Cycle-GAN代码解读

最新推荐文章于 2024-10-04 02:22:22 发布

HHzdh

最新推荐文章于 2024-10-04 02:22:22 发布

阅读量4.3k

点赞数 13

分类专栏： PyTorch对抗生成网络文章标签： pytorch python

本文链接：https://blog.csdn.net/weixin_44855366/article/details/120179781

版权

PyTorch对抗生成网络专栏收录该内容

26 篇文章 13 订阅

订阅专栏

源码：https://github.com/eriklindernoren/PyTorch-GAN/tree/master/implementations/cyclegan

1 model.py文件

1.1 初始化函数

import torch.nn as nn
import torch.nn.functional as F
import torch

# 初始化函数
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)
        if hasattr(m, "bias") and m.bias is not None:
            torch.nn.init.constant_(m.bias.data, 0.0)
    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)

1.2 RESNET 模块定义

class ResidualBlock(nn.Module):
    def __init__(self, in_features):
        super(ResidualBlock, self).__init__()

        self.block = nn.Sequential(
            nn.ReflectionPad2d(1),
            nn.Conv2d(in_features, in_features, 3),
            nn.InstanceNorm2d(in_features),
            nn.ReLU(inplace=True),
            nn.ReflectionPad2d(1),
            nn.Conv2d(in_features, in_features, 3),
            nn.InstanceNorm2d(in_features),
        )

    def forward(self, x):
        return x + self.block(x)

从生成器中截取一个resnet模块其结构如下所示。

1.3 模型定义

生成器定义：模型一上来就是3个“卷积块”，每个卷积块包含：一个2D卷积层，一个Instance Normalization层和一个ReLU。这3个“卷积块”是用来降采样的。然后是9个“残差块”，每个残差块包含2个卷积层，每个卷积层后面都有一个Instance Normalization层，第一个Instance Normalization层后面是ReLU激活函数，这些使用残差连接。然后过3个“上采样块”，每个块包含一个2D转置卷积层，1个Instance Normalization和1个ReLU激活函数。最后一层是一个2D卷积层，使用tanh作为激活函数，该层生成的形状为(256，256，3)的图像。这个Generator的输入和输出的大小是一摸一样的，都是(256，256，3)。

class GeneratorResNet(nn.Module):
    def __init__(self, input_shape, num_residual_blocks):
        super(GeneratorResNet, self).__init__()

        channels = input_shape[0]

        # Initial convolution block
        # 初始化卷积模块
        out_features = 64
        model = [
            nn.ReflectionPad2d(channels),
            nn.Conv2d(channels, out_features, 7),
            nn.InstanceNorm2d(out_features),
            nn.ReLU(inplace=True),
        ]
        in_features = out_features

        # Downsampling
        # 降采样 3个卷积模块
        for _ in range(2):
            out_features *= 2
            model += [
                nn.Conv2d(in_features, out_features, 3, stride=2, padding=1),
                nn.InstanceNorm2d(out_features),
                nn.ReLU(inplace=True),
            ]
            in_features = out_features

        # Residual blocks
        # resnet模块  num_residual_blocks=9
        for _ in range(num_residual_blocks):
            model += [ResidualBlock(out_features)]

        # Upsampling
        # 上采样
        for _ in range(2):
            out_features //= 2
            model += [
                nn.Upsample(scale_factor=2),
                nn.Conv2d(in_features, out_features, 3, stride=1, padding=1),
                nn.InstanceNorm2d(out_features),
                nn.ReLU(inplace=True),
            ]
            in_features = out_features

        # Output layer
        # 输出层
        model += [nn.ReflectionPad2d(channels), nn.Conv2d(out_features, channels, 7), nn.Tanh()]

        self.model = nn.Sequential(*model)

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

判别器定义：判别网络的架构类似于PatchGAN中的判别网络架构，是一个包含几个卷积块的深度卷积神经网络。

class Discriminator(nn.Module):
    def __init__(self, input_shape):
        super(Discriminator, self).__init__()

        channels, height, width = input_shape

        # Calculate output shape of image discriminator (PatchGAN)
        # 计算判别器输出的图片大小(PatchGAN)
        self.output_shape = (1, height // 2 ** 4, width // 2 ** 4)

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

        self.model = nn.Sequential(
            *discriminator_block(channels, 64, normalize=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)
        )

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

判别器的结构如下所示：

2 datasets.py文件

首先数据集文件是这样的：需要分别对两种风格的图片进行读取。数据下载位置（https://people.eecs.berkeley.edu/~taesung_park/CycleGAN/datasets/）。

主要是ImageDataset函数的操作，__init__操作将trainA和trainB的路径读入files_A 和files_B；__getitem__对两个文件夹的图片进行读取，若不是RGB图片则进行转换；__len__返回两个文件夹数据数量的大值。

import glob
import random
import os

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


# 转为rgb图片
def to_rgb(image):
    rgb_image = Image.new("RGB", image.size)
    rgb_image.paste(image)
    return rgb_image


# 对数据进行读取
class ImageDataset(Dataset):
    def __init__(self, root, transforms_=None, unaligned=False, mode="train"):
        self.transform = transforms.Compose(transforms_)
        self.unaligned = unaligned

        self.files_A = sorted(glob.glob(os.path.join(root, "trainA") + "/*.*"))
        self.files_B = sorted(glob.glob(os.path.join(root, "trainB") + "/*.*"))
        '''
        self.files_A = sorted(glob.glob(os.path.join(root, "%s/A" % mode) + "/*.*"))
        self.files_B = sorted(glob.glob(os.path.join(root, "%s/B" % mode) + "/*.*"))
        '''


    def __getitem__(self, index):
        image_A = Image.open(self.files_A[index % len(self.files_A)])

        if self.unaligned:
            image_B = Image.open(self.files_B[random.randint(0, len(self.files_B) - 1)])
        else:
            image_B = Image.open(self.files_B[index % len(self.files_B)])

        # Convert grayscale images to rgb
        if image_A.mode != "RGB":
            image_A = to_rgb(image_A)
        if image_B.mode != "RGB":
            image_B = to_rgb(image_B)

        item_A = self.transform(image_A)
        item_B = self.transform(image_B)
        return {"A": item_A, "B": item_B}

    def __len__(self):
        return max(len(self.files_A), len(self.files_B))

3 utils.py文件

主要关注学习率衰减（LambdaLR）。

import random
import time
import datetime
import sys

from torch.autograd import Variable
import torch
import numpy as np

from torchvision.utils import save_image


class ReplayBuffer:
    def __init__(self, max_size=50):
        assert max_size > 0, "Empty buffer or trying to create a black hole. Be careful."
        self.max_size = max_size
        self.data = []

    def push_and_pop(self, data):
        to_return = []
        for element in data.data:
            element = torch.unsqueeze(element, 0)
            if len(self.data) < self.max_size:
                self.data.append(element)
                to_return.append(element)
            else:
                if random.uniform(0, 1) > 0.5:
                    i = random.randint(0, self.max_size - 1)
                    to_return.append(self.data[i].clone())
                    self.data[i] = element
                else:
                    to_return.append(element)
        return Variable(torch.cat(to_return))


class LambdaLR:
    def __init__(self, n_epochs, offset, decay_start_epoch):
        assert (n_epochs - decay_start_epoch) > 0, "Decay must start before the training session ends!"
        self.n_epochs = n_epochs
        self.offset = offset
        self.decay_start_epoch = decay_start_epoch

    def step(self, epoch):
        return 1.0 - max(0, epoch + self.offset - self.decay_start_epoch) / (self.n_epochs - self.decay_start_epoch)

4 cyclegan.py文件

4.1 导入相关库以及进行参数设置

导入相关库

import argparse
import os
import numpy as np
import math
import itertools
import datetime
import time

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

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

from models import *
from datasets import *
from utils import *

import torch.nn as nn
import torch.nn.functional as F
import torch

参数设置

'''参数表格
epoch:使用数据集的所有数据进行一次模型训练，一代训练，从第0代开始训练
n_epochs:训练的次数，默认200次
dataset_name：数据集文件夹的名字，默认"monet2photo"
batch_size：使用数据中的一部分数据进行模型权重更新的这部分数据大小，默认1,(受限于电脑性能)
lr：adam学习率
b1&b2：adam学习参数
decay_epoch：lr学习率开始衰减
n_cpu：训练过程中用到的CPU线程数目
img_height：输入图片的高度，默认256
img_width：输入图片的宽度，默认256
channels：图片的通道数，默认为彩色图片，channels=3
sample_interval：每隔一段时间对训练输出进行采样并展示，默认100
n_residual_blocks：生成器中的residual模块的数量
lambda_cyc：cycle loss权重参数
lambda_id：identity loss权重参数
'''

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="monet2photo", 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=8, help="number of cpu threads to use during batch generation")
parser.add_argument("--img_height", type=int, default=256, help="size of image height")
parser.add_argument("--img_width", type=int, default=256, help="size of image width")
parser.add_argument("--channels", type=int, default=3, help="number of image channels")
parser.add_argument("--sample_interval", type=int, default=100, help="interval between saving generator outputs")
parser.add_argument("--checkpoint_interval", type=int, default=-1, help="interval between saving model checkpoints")
parser.add_argument("--n_residual_blocks", type=int, default=9, help="number of residual blocks in generator")
parser.add_argument("--lambda_cyc", type=float, default=10.0, help="cycle loss weight")
parser.add_argument("--lambda_id", type=float, default=5.0, help="identity loss weight")
opt = parser.parse_args()
print(opt)

得到参数opt为：

创建文件夹保存模型和采样输出图片

# 创建文件夹来保存训练过程的采样输出以及保存模型
# Create sample and checkpoint directories
os.makedirs("images/%s" % opt.dataset_name, exist_ok=True)
os.makedirs("saved_models/%s" % opt.dataset_name, exist_ok=True)

4.2 损失函数定义和初始化

# 初始化三个损失函数
# Losses
criterion_GAN = torch.nn.MSELoss()
criterion_cycle = torch.nn.L1Loss()
criterion_identity = torch.nn.L1Loss()

4.3 初始化生成器、判别器

# 判断电脑是否可以使用GPU进行训练
cuda = torch.cuda.is_available()

# input_shape保存输入图片的通道数，高度，宽度
input_shape = (opt.channels, opt.img_height, opt.img_width)

# 初始化四个网络(G_AB,G_BA,D_A,D_B)
# Initialize generator and discriminator
G_AB = GeneratorResNet(input_shape, opt.n_residual_blocks)
G_BA = GeneratorResNet(input_shape, opt.n_residual_blocks)
D_A = Discriminator(input_shape)
D_B = Discriminator(input_shape)

4.4 导入参数

# 采用GPU进行训练
if cuda:
    G_AB = G_AB.cuda()
    G_BA = G_BA.cuda()
    D_A = D_A.cuda()
    D_B = D_B.cuda()
    criterion_GAN.cuda()
    criterion_cycle.cuda()
    criterion_identity.cuda()

# 如果不是从第0代开始训练，则从保存的模型中调用模型以及加载开始训练的代数，继续训练
if opt.epoch != 0:
    # Load pretrained models
    G_AB.load_state_dict(torch.load("saved_models/%s/G_AB_%d.pth" % (opt.dataset_name, opt.epoch)))
    G_BA.load_state_dict(torch.load("saved_models/%s/G_BA_%d.pth" % (opt.dataset_name, opt.epoch)))
    D_A.load_state_dict(torch.load("saved_models/%s/D_A_%d.pth" % (opt.dataset_name, opt.epoch)))
    D_B.load_state_dict(torch.load("saved_models/%s/D_B_%d.pth" % (opt.dataset_name, opt.epoch)))
else:
    # Initialize weights
    G_AB.apply(weights_init_normal)
    G_BA.apply(weights_init_normal)
    D_A.apply(weights_init_normal)
    D_B.apply(weights_init_normal)

4.5 优化器设置

# 定义初始化模型的优化器
# Optimizers
optimizer_G = torch.optim.Adam(
    itertools.chain(G_AB.parameters(), G_BA.parameters()), lr=opt.lr, betas=(opt.b1, opt.b2)
)
optimizer_D_A = torch.optim.Adam(D_A.parameters(), lr=opt.lr, betas=(opt.b1, opt.b2))
optimizer_D_B = torch.optim.Adam(D_B.parameters(), lr=opt.lr, betas=(opt.b1, opt.b2))

4.6 学习率衰减

torch.optim.lr_scheduler.LambdaLR里面有一个参数lr_lambda是要输入学习率，这个学习率是动态变化的，由LambdaLR(opt.n_epochs, opt.epoch, opt.decay_epoch).step来获得。

# 按照epoch的次数自动调整学习率
# Learning rate update schedulers
lr_scheduler_G = torch.optim.lr_scheduler.LambdaLR(
    optimizer_G, lr_lambda=LambdaLR(opt.n_epochs, opt.epoch, opt.decay_epoch).step
)
lr_scheduler_D_A = torch.optim.lr_scheduler.LambdaLR(
    optimizer_D_A, lr_lambda=LambdaLR(opt.n_epochs, opt.epoch, opt.decay_epoch).step
)
lr_scheduler_D_B = torch.optim.lr_scheduler.LambdaLR(
    optimizer_D_B, lr_lambda=LambdaLR(opt.n_epochs, opt.epoch, opt.decay_epoch).step
)

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

4.7 数据读取和数据预处理

数据预处理包括resize、crop、flip、normalize等操作。

# Buffers of previously generated samples
fake_A_buffer = ReplayBuffer()
fake_B_buffer = ReplayBuffer()

# 图片预处理函数
# Image transformations
transforms_ = [
    transforms.Resize(int(opt.img_height * 1.12), Image.BICUBIC),  # 调整Image对象的尺寸
    transforms.RandomCrop((opt.img_height, opt.img_width)),  # 扩大后剪切成img_height*img_width大小的图片
    transforms.RandomHorizontalFlip(),  # 依据概率p对PIL图片进行水平翻转，p默认0.5
    transforms.ToTensor(),  # 转为tensor格式
    transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)),  # 归一化
]


#print("../../data/%s" % opt.dataset_name)
#test = ImageDataset("E:\python\pytorch-gan-master\data\monet2photo" , transforms_=transforms_, unaligned=True)

# 加载训练数据
# Training data loader
dataloader = DataLoader(
    # ../表示当前目录的父目录
    ImageDataset("../../data/%s" % opt.dataset_name, transforms_=transforms_, unaligned=True),
    batch_size=opt.batch_size,
    shuffle=True,  # 将数据打乱，数值越大，混乱程度越大
    # num_workers=0,
    num_workers=opt.n_cpu,  # 线程数
)

# 测试数据加载
# Test data loader
val_dataloader = DataLoader(
    ImageDataset("../../data/%s" % opt.dataset_name, transforms_=transforms_, unaligned=True, mode="test"),
    batch_size=5,
    shuffle=True,
    num_workers=0,
)

4.8 保存训练结果的样本

# 定义测试数据喂进网络的输出展示函数
def sample_images(batches_done):
    """Saves a generated sample from the test set"""
    imgs = next(iter(val_dataloader))
    G_AB.eval()
    G_BA.eval()
    real_A = Variable(imgs["A"].type(Tensor))
    fake_B = G_AB(real_A)
    real_B = Variable(imgs["B"].type(Tensor))
    fake_A = G_BA(real_B)
    # Arange images along x-axis
    real_A = make_grid(real_A, nrow=5, normalize=True)
    real_B = make_grid(real_B, nrow=5, normalize=True)
    fake_A = make_grid(fake_A, nrow=5, normalize=True)
    fake_B = make_grid(fake_B, nrow=5, normalize=True)
    # Arange images along y-axis
    image_grid = torch.cat((real_A, fake_B, real_B, fake_A), 1)
    save_image(image_grid, "images/%s/%s.png" % (opt.dataset_name, batches_done), normalize=False)

4.9 开始训练和保存模型

# ----------
#  Training
#  开始训练
# ----------

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

        # 设置模型输入
        # Set model input
        real_A = Variable(batch["A"].type(Tensor))
        real_B = Variable(batch["B"].type(Tensor))

        # 对抗生成网络中的真实图片和虚假图片
        # Adversarial ground truths
        valid = Variable(Tensor(np.ones((real_A.size(0), *D_A.output_shape))), requires_grad=False)
        fake = Variable(Tensor(np.zeros((real_A.size(0), *D_A.output_shape))), requires_grad=False)

        # ------------------
        #  Train Generators
        #  训练生成器
        # ------------------

        G_AB.train()
        G_BA.train()

        # 梯度清零，方便下代训练
        optimizer_G.zero_grad()

        # Identity loss ：用于保证生成图像的连续性，一个图像x，经过其中一个生成器生成图像 G(x)，尽可能与原来图像接近。
        loss_id_A = criterion_identity(G_BA(real_A), real_A)
        loss_id_B = criterion_identity(G_AB(real_B), real_B)

        loss_identity = (loss_id_A + loss_id_B) / 2

        # GAN loss
        fake_B = G_AB(real_A)
        loss_GAN_AB = criterion_GAN(D_B(fake_B), valid)
        fake_A = G_BA(real_B)
        loss_GAN_BA = criterion_GAN(D_A(fake_A), valid)

        loss_GAN = (loss_GAN_AB + loss_GAN_BA) / 2

        # Cycle loss
        recov_A = G_BA(fake_B)
        loss_cycle_A = criterion_cycle(recov_A, real_A)
        recov_B = G_AB(fake_A)
        loss_cycle_B = criterion_cycle(recov_B, real_B)

        loss_cycle = (loss_cycle_A + loss_cycle_B) / 2

        # 总损失函数
        # Total loss
        loss_G = loss_GAN + opt.lambda_cyc * loss_cycle + opt.lambda_id * loss_identity

        # 反向传播
        loss_G.backward()

        # 权重更新
        optimizer_G.step()

        # -----------------------
        #  Train Discriminator A
        #  训练分类器A
        # -----------------------

        optimizer_D_A.zero_grad()

        # Real loss
        loss_real = criterion_GAN(D_A(real_A), valid)
        # Fake loss (on batch of previously generated samples)
        fake_A_ = fake_A_buffer.push_and_pop(fake_A)
        loss_fake = criterion_GAN(D_A(fake_A_.detach()), fake)
        # Total loss
        loss_D_A = (loss_real + loss_fake) / 2

        loss_D_A.backward()
        optimizer_D_A.step()

        # -----------------------
        #  Train Discriminator B
        #  训练分类器B
        # -----------------------

        optimizer_D_B.zero_grad()

        # Real loss
        loss_real = criterion_GAN(D_B(real_B), valid)
        # Fake loss (on batch of previously generated samples)
        fake_B_ = fake_B_buffer.push_and_pop(fake_B)
        loss_fake = criterion_GAN(D_B(fake_B_.detach()), fake)
        # Total loss
        loss_D_B = (loss_real + loss_fake) / 2

        loss_D_B.backward()
        optimizer_D_B.step()

        loss_D = (loss_D_A + loss_D_B) / 2

        # --------------
        #  Log Progress
        #  训练进程显示
        # --------------

        # 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, adv: %f, cycle: %f, identity: %f] ETA: %s"
            % (
                epoch,
                opt.n_epochs,
                i,
                len(dataloader),
                loss_D.item(),
                loss_G.item(),
                loss_GAN.item(),
                loss_cycle.item(),
                loss_identity.item(),
                time_left,
            )
        )

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

    # Update learning rates
    # 更新学习率
    lr_scheduler_G.step()
    lr_scheduler_D_A.step()
    lr_scheduler_D_B.step()

    if opt.checkpoint_interval != -1 and epoch % opt.checkpoint_interval == 0:
        # Save model checkpoints
        torch.save(G_AB.state_dict(), "saved_models/%s/G_AB_%d.pth" % (opt.dataset_name, epoch))
        torch.save(G_BA.state_dict(), "saved_models/%s/G_BA_%d.pth" % (opt.dataset_name, epoch))
        torch.save(D_A.state_dict(), "saved_models/%s/D_A_%d.pth" % (opt.dataset_name, epoch))
        torch.save(D_B.state_dict(), "saved_models/%s/D_B_%d.pth" % (opt.dataset_name, epoch))