【延伸阅读】让老照片重现光彩（五）：Pix2PixHD模型源代码+中文注释

英伟达公司和加州大学伯克利分校于2018年发表的“基于有条件GAN的高分辨率图像合成及语义操控”项目，是本项目“让老照片重现光彩”的技术基础，算是一个前置开源项目。

“基于有条件GAN的高分辨率图像合成及语义操控”项目的技术核心是Pix2PixHD模型，我们在这里分享一下相关的源代码+中文注释，基于此可以加深对“让老照片重现光彩”项目的理解（尤其是，在老照片项目的模型与训练源代码尚未开源的情况下）。

“基于有条件GAN的高分辨率图像合成及语义操控”项目在GitHub上的链接是：https://github.com/NVIDIA/pix2pixHD

Pix2PixHD模型使用PyTorch构建，代码清晰、整齐，相关的源代码主要是3个文件，分别是：./models/models.py、 ./models/pix2pixHD_model.py 和  ./models/networks.py

说明如下：

（1）./models/models.py

调用 Pix2PixHDModel() 创建模型。

import torch

# 创建模型，并返回模型
def create_model(opt):
    if opt.model == 'pix2pixHD':  # 选择pix2pixHD model
        from .pix2pixHD_model import Pix2PixHDModel, InferenceModel
        if opt.isTrain:  # 若是训练，则为True
            model = Pix2PixHDModel()
        else:  # 否则，若仅仅是前向传播用来演示，则为False
            model = InferenceModel()
    else:  # 选择 UIModel model
    	from .ui_model import UIModel
    	model = UIModel()
    model.initialize(opt)  # 模型初始化参数
    if opt.verbose:  # 默认为false,表示之前并无模型保存
        print("model [%s] was created" % (model.name()))  # 打印label2city模型被创建

    if opt.isTrain and len(opt.gpu_ids) and not opt.fp16:
        model = torch.nn.DataParallel(model, device_ids=opt.gpu_ids)  # 多GPU训练

    return model

（2）./models/pix2pixHD_model.py 

构建模型的核心内容：

定义有条件GAN（Pix2PixHDModel）的生成器、鉴别器、编码器（用于生成实例的低维特征）；

定义损失函数（包括：GANloss，VGGloss、特征匹配损失函数）；

定义生成器和鉴别器的优化器（optimizer）；

定义各模块的输入；

定义forward函数。

import numpy as np
import torch
import os
from torch.autograd import Variable
from util.image_pool import ImagePool
from .base_model import BaseModel
from . import networks

class Pix2PixHDModel(BaseModel):
    def name(self):
        return 'Pix2PixHDModel'

    # loss滤波器：其中g_gan、d_real、d_fake三个loss值是肯定返回的
    # 这里的g_gan_feat即论文中的“特征匹配损失函数”（论文中的等式（4））
    # g_vgg为论文中的VGG感知损失函数，稍微改善了输出结果
    # g_gan_feat、g_vgg两个loss值根据train_options的opt.no_ganFeat_loss, not opt.no_vgg_loss而定（默认是需要返回的）
    def init_loss_filter(self, use_gan_feat_loss, use_vgg_loss):
        flags = (True, use_gan_feat_loss, use_vgg_loss, True, True)
        def loss_filter(g_gan, g_gan_feat, g_vgg, d_real, d_fake):
            return [l for (l,f) in zip((g_gan,g_gan_feat,g_vgg,d_real,d_fake),flags) if f]
        return loss_filter
    
    def initialize(self, opt):
        BaseModel.initialize(self, opt)
        if opt.resize_or_crop != 'none' or not opt.isTrain: # when training at full res this causes OOM
            torch.backends.cudnn.benchmark = True
        self.isTrain = opt.isTrain
        self.use_features = opt.instance_feat or opt.label_feat
        self.gen_features = self.use_features and not self.opt.load_features
        input_nc = opt.label_nc if opt.label_nc != 0 else opt.input_nc

        ##### define networks        
        # Generator network
        # 生成器网络
        netG_input_nc = input_nc        
        if not opt.no_instance:
            netG_input_nc += 1  # 添加instance通道（区分不同实例）
        if self.use_features:
            netG_input_nc += opt.feat_num  # 添加feature_map通道（使用encoder）
        self.netG = networks.define_G(netG_input_nc, opt.output_nc, opt.ngf, opt.netG, 
                                      opt.n_downsample_global, opt.n_blocks_global, opt.n_local_enhancers, 
                                      opt.n_blocks_local, opt.norm, gpu_ids=self.gpu_ids)        

        # Discriminator network
        # 鉴别器网络
        if self.isTrain:
            use_sigmoid = opt.no_lsgan
            netD_input_nc = input_nc + opt.output_nc  # real_images + fake_images
            if not opt.no_instance:
                netD_input_nc += 1  # 添加instance通道（区分不同实例）
            self.netD = networks.define_D(netD_input_nc, opt.ndf, opt.n_layers_D, opt.norm, use_sigmoid, 
                                          opt.num_D, not opt.no_ganFeat_loss, gpu_ids=self.gpu_ids)

        ### Encoder network
        # 编码器网络（是define_G()中的一个子函数）
        if self.gen_features:          
            self.netE = networks.define_G(opt.output_nc, opt.feat_num, opt.nef, 'encoder', 
                                          opt.n_downsample_E, norm=opt.norm, gpu_ids=self.gpu_ids)  
        if self.opt.verbose:
                print('---------- Networks initialized -------------')

        # load networks
        # 加载网络（模型）
        if not self.isTrain or opt.continue_train or opt.load_pretrain:
            pretrained_path = '' if not self.isTrain else opt.load_pretrain
            self.load_network(self.netG, 'G', opt.which_epoch, pretrained_path)            
            if self.isTrain:
                self.load_network(self.netD, 'D', opt.which_epoch, pretrained_path)  
            if self.gen_features:
                self.load_network(self.netE, 'E', opt.which_epoch, pretrained_path)              

        # set loss functions and optimizers
        if self.isTrain:
            if opt.pool_size > 0 and (len(self.gpu_ids)) > 1:
                raise NotImplementedError("Fake Pool Not Implemented for MultiGPU")
            self.fake_pool = ImagePool(opt.pool_size)  # 初始化fake_pool：num_imgs = 0，images = []
            self.old_lr = opt.lr

            # define loss functions
            # 定义损失函数，在.forward()中使用
            # 默认使用ganfeat_loss和vgg_loss
            self.loss_filter = self.init_loss_filter(not opt.no_ganFeat_loss, not opt.no_vgg_loss)
            
            self.criterionGAN = networks.GANLoss(use_lsgan=not opt.no_lsgan, tensor=self.Tensor)   
            self.criterionFeat = torch.nn.L1Loss()
            if not opt.no_vgg_loss:             
                self.criterionVGG = networks.VGGLoss(self.gpu_ids)
                
            # Names so we can breakout loss
            # 给损失函数命名
            self.loss_names = self.loss_filter('G_GAN','G_GAN_Feat','G_VGG','D_real', 'D_fake')

            # initialize optimizers
            # 初始化优化器
            # optimizer G（含：encoder）
            if opt.niter_fix_global > 0:                
                import sys
                if sys.version_info >= (3,0):
                    finetune_list = set()
                else:
                    from sets import Set
                    finetune_list = Set()

                params_dict = dict(self.netG.named_parameters())
                params = []
                for key, value in params_dict.items():       
                    if key.startswith('model' + str(opt.n_local_enhancers)):                    
                        params += [value]
                        finetune_list.add(key.split('.')[0])  
                print('------------- Only training the local enhancer network (for %d epochs) ------------' % opt.niter_fix_global)
                print('The layers that are finetuned are ', sorted(finetune_list))                         
            else:
                params = list(self.netG.parameters())
            if self.gen_features:              
                params += list(self.netE.parameters())         
            self.optimizer_G = torch.optim.Adam(params, lr=opt.lr, betas=(opt.beta1, 0.999))                            

            # optimizer D                        
            params = list(self.netD.parameters())    
            self.optimizer_D = torch.optim.Adam(params, lr=opt.lr, betas=(opt.beta1, 0.999))

    # feat=feature（特征），inst=instance（实例）
    # label_map（标签图）每个像素值代表像素的对象类，inst_map（实例图）每个像素包含每个单独对象的唯一对象ID
    # 获取实例图的边界（边缘），将edge_map与label_map的one-hot向量拼接在一起，封装为Variable，赋值给input_label
    # real_image和feat_map，封装为Variable，赋值给real_image和feat_map；label_map赋值给inst_map
    def encode_input(self, label_map, inst_map=None, real_image=None, feat_map=None, infer=False):
        # label_map 数据类型转化
        if self.opt.label_nc == 0:
            input_label = label_map.data.cuda()
        else:
            # create one-hot vector for label map 
            size = label_map.size()
            oneHot_size = (size[0], self.opt.label_nc, size[2], size[3])
            input_label = torch.cuda.FloatTensor(torch.Size(oneHot_size)).zero_()
            input_label = input_label.scatter_(1, label_map.data.long().cuda(), 1.0)  # 将列表转成one-hot编码的形式
            if self.opt.data_type == 16:
                input_label = input_label.half()

        # get edges from instance map
        # 获取实例图的边界（边缘），将edge_map与input_label拼接在一起
        if not self.opt.no_instance:
            inst_map = inst_map.data.cuda()
            edge_map = self.get_edges(inst_map)
            input_label = torch.cat((input_label, edge_map), dim=1)         
        input_label = Variable(input_label, volatile=infer)

        # real images for training
        if real_image is not None:
            real_image = Variable(real_image.data.cuda())

        # instance map for feature encoding
        if self.use_features:
            # get precomputed feature maps
            if self.opt.load_features:
                feat_map = Variable(feat_map.data.cuda())
            if self.opt.label_feat:
                inst_map = label_map.cuda()

        return input_label, inst_map, real_image, feat_map

    # 定义判别器
    def discriminate(self, input_label, test_image, use_pool=False):
        input_concat = torch.cat((input_label, test_image.detach()), dim=1)
        if use_pool:            
            fake_query = self.fake_pool.query(input_concat)  # 读取fake images
            return self.netD.forward(fake_query)
        else:
            return self.netD.forward(input_concat)

    # 前向传播，使用输入数据运行模型
    # PyTorch 允许在前向传播过程中进行动态操作（如：跳跃连接等）
    def forward(self, label, inst, image, feat, infer=False):
        # Encode Inputs
        # 获取实例图的边界（边缘），将edge_map与label_map的one-hot向量拼接在一起，封装为Variable，赋值给input_label
        input_label, inst_map, real_image, feat_map = self.encode_input(label, inst, image, feat)  

        # Fake Generation
        # 调用生成器生成fake images
        if self.use_features:
            # 调用netE（即：encoder）对输入图片进行encoder-decoder运算，提取feature_map
            if not self.opt.load_features:
                feat_map = self.netE.forward(real_image, inst_map)                     
            input_concat = torch.cat((input_label, feat_map), dim=1)  # 将input_label与特征图拼接在一起，作为生成器netG的输入
        else:
            input_concat = input_label
        fake_image = self.netG.forward(input_concat)

        # Fake Detection and Loss
        # 输入为input_label和fake_image，鉴别器生成fake images pool（假图片池）的预测（prediction）、D_fake损失函数
        pred_fake_pool = self.discriminate(input_label, fake_image, use_pool=True)
        loss_D_fake = self.criterionGAN(pred_fake_pool, False)

        # Real Detection and Loss
        # 输入为input_label和real_image，鉴别器生成real images的预测（prediction）、D_real损失函数
        pred_real = self.discriminate(input_label, real_image)
        loss_D_real = self.criterionGAN(pred_real, True)

        # GAN loss (Fake Passability Loss)
        # 将输入标签与假图片拼接后作为输入，鉴别器生成假图片预测（prediction）、G_GAN损失函数
        pred_fake = self.netD.forward(torch.cat((input_label, fake_image), dim=1))        
        loss_G_GAN = self.criterionGAN(pred_fake, True)

        # GAN feature matching loss
        # 计算GAN的特征匹配损失函数，每一个尺度的鉴别器（num_D）、鉴别器的每层特征提取器（pred_fake）分别加权计算并求和
        loss_G_GAN_Feat = 0
        if not self.opt.no_ganFeat_loss:
            feat_weights = 4.0 / (self.opt.n_layers_D + 1)  # 4.0/(鉴别器的层数+1)
            D_weights = 1.0 / self.opt.num_D  # 1.0/(多尺度的个数，论文中是3）
            for i in range(self.opt.num_D):
                for j in range(len(pred_fake[i])-1):
                    # 计算：L1Loss()，lambda_feat为（输入的）调节系数
                    loss_G_GAN_Feat += D_weights * feat_weights * \
                        self.criterionFeat(pred_fake[i][j], pred_real[i][j].detach()) * self.opt.lambda_feat
                   
        # VGG feature matching loss
        # VGG特征匹配损失函数
        loss_G_VGG = 0
        if not self.opt.no_vgg_loss:
            # 计算fake_image和real_image之间的VGGLoss，lambda_feat为输入的调节系数
            # real_image不进行梯度计算
            loss_G_VGG = self.criterionVGG(fake_image, real_image) * self.opt.lambda_feat
        
        # Only return the fake_B image if necessary to save BW
        return [ self.loss_filter( loss_G_GAN, loss_G_GAN_Feat, loss_G_VGG, loss_D_real, loss_D_fake ), None if not infer else fake_image ]

    # 推理
    # 将标签、实例边界、特征图作为输入，生成假图片
    def inference(self, label, inst, image=None):
        # Encode Inputs        
        image = Variable(image) if image is not None else None
        # 将实例边界与label的one-hot向量拼接在一起，返回给input_label
        input_label, inst_map, real_image, _ = self.encode_input(Variable(label), Variable(inst), image, infer=True)

        # Fake Generation
        if self.use_features:
            if self.opt.use_encoded_image:
                # encode the real image to get feature map
                # 用encoder计算真实图像的特征图
                feat_map = self.netE.forward(real_image, inst_map)
            else:
                # sample clusters from precomputed features
                # 随机选取实例图中的某个特征作为编码特征，用于训练
                feat_map = self.sample_features(inst_map)
            input_concat = torch.cat((input_label, feat_map), dim=1)  # 把feat_map和input_label拼接在一起，作为生成器的输入
        else:
            input_concat = input_label        
           
        if torch.__version__.startswith('0.4'):
            with torch.no_grad():
                fake_image = self.netG.forward(input_concat)  # 调用generator生成假图片
        else:
            fake_image = self.netG.forward(input_concat)
        return fake_image

    def sample_features(self, inst): 
        # read precomputed feature clusters 
        cluster_path = os.path.join(self.opt.checkpoints_dir, self.opt.name, self.opt.cluster_path)        
        features_clustered = np.load(cluster_path, encoding='latin1').item()

        # randomly sample from the feature clusters
        # 从特征簇中随机采样
        inst_np = inst.cpu().numpy().astype(int)                                      
        feat_map = self.Tensor(inst.size()[0], self.opt.feat_num, inst.size()[2], inst.size()[3])  # feat_map.size
        for i in np.unique(inst_np):  # 对于一维数组或者列表，unique()去除其中重复的元素，并按元素由大到小返回一个新的无元素重复的元组或者列表
                                      # 确定具有唯一性的特征代码，并将特征代码排序
            label = i if i < 1000 else i//1000
            if label in features_clustered:
                feat = features_clustered[label]  # 从特征簇中取出当前特征代码对应的特征向量
                cluster_idx = np.random.randint(0, feat.shape[0])   # 任取一个随机数，用于抽取feat[]的某一行数据
                                            
                idx = (inst == int(i)).nonzero()  # nonzero()返回非零的位置，即特征图中与排序后的特征代码一致的所有非零位置
                for k in range(self.opt.feat_num):  # feat_num，特征的个数
                    # feat_map[channel, feature_num, hight, width]
                    # 任意抽取feat[]中某一行中的数据，赋值给feat_map
                    feat_map[idx[:,0], idx[:,1] + k, idx[:,2], idx[:,3]] = feat[cluster_idx, k]
        if self.opt.data_type==16:
            feat_map = feat_map.half()
        return feat_map

    def encode_features(self, image, inst):
        image = Variable(image.cuda(), volatile=True)
        feat_num = self.opt.feat_num
        h, w = inst.size()[2], inst.size()[3]
        block_num = 32
        feat_map = self.netE.forward(image, inst.cuda())
        inst_np = inst.cpu().numpy().astype(int)
        feature = {}
        for i in range(self.opt.label_nc):
            feature[i] = np.zeros((0, feat_num+1))
        for i in np.unique(inst_np):
            label = i if i < 1000 else i//1000
            idx = (inst == int(i)).nonzero()
            num = idx.size()[0]
            idx = idx[num//2,:]
            val = np.zeros((1, feat_num+1))                        
            for k in range(feat_num):
                val[0, k] = feat_map[idx[0], idx[1] + k, idx[2], idx[3]].data[0]            
            val[0, feat_num] = float(num) / (h * w // block_num)
            feature[label] = np.append(feature[label], val, axis=0)
        return feature

    # 获得instance的边界（边缘），t是inst_map
    # 如果实例边界图中的一个像素的对象ID与它的4个邻居中的任何一个不同，那么该像素为1，否则为0
    def get_edges(self, t):
        edge = torch.cuda.ByteTensor(t.size()).zero_()  # 初始化为0
        edge[:,:,:,1:] = edge[:,:,:,1:] | (t[:,:,:,1:] != t[:,:,:,:-1])
        edge[:,:,:,:-1] = edge[:,:,:,:-1] | (t[:,:,:,1:] != t[:,:,:,:-1])
        edge[:,:,1:,:] = edge[:,:,1:,:] | (t[:,:,1:,:] != t[:,:,:-1,:])
        edge[:,:,:-1,:] = edge[:,:,:-1,:] | (t[:,:,1:,:] != t[:,:,:-1,:])
        if self.opt.data_type==16:
            return edge.half()
        else:
            return edge.float()

    # 保存模型参数
    def save(self, which_epoch):
        self.save_network(self.netG, 'G', which_epoch, self.gpu_ids)
        self.save_network(self.netD, 'D', which_epoch, self.gpu_ids)
        if self.gen_features:
            self.save_network(self.netE, 'E', which_epoch, self.gpu_ids)

    def update_fixed_params(self):
        # after fixing the global generator for a number of iterations, also start finetuning it
        params = list(self.netG.parameters())
        if self.gen_features:
            params += list(self.netE.parameters())           
        self.optimizer_G = torch.optim.Adam(params, lr=self.opt.lr, betas=(self.opt.beta1, 0.999))
        if self.opt.verbose:
            print('------------ Now also finetuning global generator -----------')

    # 更新学习率
    def update_learning_rate(self):
        lrd = self.opt.lr / self.opt.niter_decay
        lr = self.old_lr - lrd        
        for param_group in self.optimizer_D.param_groups:
            param_group['lr'] = lr
        for param_group in self.optimizer_G.param_groups:
            param_group['lr'] = lr
        if self.opt.verbose:
            print('update learning rate: %f -> %f' % (self.old_lr, lr))
        self.old_lr = lr

# 推理模型，前向传播
class InferenceModel(Pix2PixHDModel):
    def forward(self, inp):
        label, inst = inp
        return self.inference(label, inst)

        

（3） ./models/networks.py

定义底层的神经网络模块：

定义生成器define_G()，以及生成器中的核心模块：全局生成器GlobalGenerator()、局部增强器LocalEnhancer()、残差块ResnetBlock()、编码器Encoder()；

定义鉴别器define_D()，以及鉴别器的核心模块：多尺度鉴别器MultiscaleDiscriminator()、PactchGAN N层鉴别器NLayerDiscriminator()；

定义损失函数GANLoss()、VGGLoss()；

定义网络模型Vgg19()。

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

###############################################################################
# Functions
###############################################################################
def weights_init(m):
    classname = m.__class__.__name__
    if classname.find('Conv') != -1:
        m.weight.data.normal_(0.0, 0.02)
    elif classname.find('BatchNorm2d') != -1:
        m.weight.data.normal_(1.0, 0.02)
        m.bias.data.fill_(0)

# 数据的归一化处理
def get_norm_layer(norm_type='instance'):
    if norm_type == 'batch':
        norm_layer = functools.partial(nn.BatchNorm2d, affine=True  # 对NHW做归一化
    elif norm_type == 'instance':
        norm_layer = functools.partial(nn.InstanceNorm2d, affine=False)  # 对HW做归一化，用在风格化迁移
    else:
        raise NotImplementedError('normalization layer [%s] is not found' % norm_type)
    return norm_layer

# 在Pix2PixHD中，G分为两部分，一部分是global net，另一部分是local net，即：define_G()前两个if语句对应的分支
# 第三个if语句对应的是论文中E的部分，用来预先计算类别特征，区分相同语义标签（semantic label）的多个实例
# input_nc = 3，number of input channels（不含instance和feature map通道）
# output_nc = 3，number of output channels（不含instance和feature map通道）
# ngf = 64 第一层卷积核数
def define_G(input_nc, output_nc, ngf, netG, n_downsample_global=3, n_blocks_global=9, n_local_enhancers=1, 
             n_blocks_local=3, norm='instance', gpu_ids=[]):    
    norm_layer = get_norm_layer(norm_type=norm)     
    if netG == 'global':    
        netG = GlobalGenerator(input_nc, output_nc, ngf, n_downsample_global, n_blocks_global, norm_layer)       
    elif netG == 'local':        
        netG = LocalEnhancer(input_nc, output_nc, ngf, n_downsample_global, n_blocks_global, 
                                  n_local_enhancers, n_blocks_local, norm_layer)
    elif netG == 'encoder':
        netG = Encoder(input_nc, output_nc, ngf, n_downsample_global, norm_layer)
    else:
        raise('generator not implemented!')
    print(netG)
    if len(gpu_ids) > 0:
        assert(torch.cuda.is_available())
        netG.cuda(gpu_ids[0])
    netG.apply(weights_init)
    return netG

# 按照论文的说法，Pix2PixHD的D有多个（3个）
# input_nc = 3+3 (real_images+fake_images，不含instance通道）
def define_D(input_nc, ndf, n_layers_D, norm='instance', use_sigmoid=False, num_D=1, getIntermFeat=False, gpu_ids=[]):        
    norm_layer = get_norm_layer(norm_type=norm)   
    netD = MultiscaleDiscriminator(input_nc, ndf, n_layers_D, norm_layer, use_sigmoid, num_D, getIntermFeat)   
    print(netD)
    if len(gpu_ids) > 0:
        assert(torch.cuda.is_available())
        netD.cuda(gpu_ids[0])
    netD.apply(weights_init)
    return netD

def print_network(net):
    if isinstance(net, list):
        net = net[0]
    num_params = 0
    for param in net.parameters():
        num_params += param.numel()
    print(net)
    print('Total number of parameters: %d' % num_params)

##############################################################################
# Losses
##############################################################################
class GANLoss(nn.Module):
    def __init__(self, use_lsgan=True, target_real_label=1.0, target_fake_label=0.0,
                 tensor=torch.FloatTensor):
        super(GANLoss, self).__init__()
        self.real_label = target_real_label
        self.fake_label = target_fake_label
        self.real_label_var = None
        self.fake_label_var = None
        self.Tensor = tensor
        # lsgan: Least Squares GAN, 最小二乘GAN
        if use_lsgan:
            self.loss = nn.MSELoss()  #  均方差 MSE(Mean Square Error)
        else:
            self.loss = nn.BCELoss()  # 二元交叉熵 BCE(Binary Cross Entropy)，xlog(p(x)) + (1-x)log(1-p(x))

    # Pytorch中基本的变量类型是FloatTensor
    # Variable是FloatTensor的封装，除了包含FloatTensor还包含有梯度信息
    def get_target_tensor(self, input, target_is_real):
        target_tensor = None
        if target_is_real:
            create_label = ((self.real_label_var is None) or
                            (self.real_label_var.numel() != input.numel()))
            if create_label:
                real_tensor = self.Tensor(input.size()).fill_(self.real_label)
                self.real_label_var = Variable(real_tensor, requires_grad=False)
            target_tensor = self.real_label_var
        else:
            create_label = ((self.fake_label_var is None) or
                            (self.fake_label_var.numel() != input.numel()))
            if create_label:
                fake_tensor = self.Tensor(input.size()).fill_(self.fake_label)
                self.fake_label_var = Variable(fake_tensor, requires_grad=False)
            target_tensor = self.fake_label_var
        return target_tensor

    def __call__(self, input, target_is_real):
        if isinstance(input[0], list):
            loss = 0
            for input_i in input:
                pred = input_i[-1]
                target_tensor = self.get_target_tensor(pred, target_is_real)
                loss += self.loss(pred, target_tensor)
            return loss
        else:            
            target_tensor = self.get_target_tensor(input[-1], target_is_real)
            return self.loss(input[-1], target_tensor)

# VGG19输出的特征图的5个切片的L1Loss()，权重分别为[1/32, 1/16, 1/8, 1/4, 1]，加权求和
class VGGLoss(nn.Module):
    def __init__(self, gpu_ids):
        super(VGGLoss, self).__init__()        
        self.vgg = Vgg19().cuda()
        self.criterion = nn.L1Loss()  # L1Loss, 平均绝对误差（Mean Absolute Error,MAE)
        self.weights = [1.0/32, 1.0/16, 1.0/8, 1.0/4, 1.0]        

    # 计算 x 和 y 的 L1Loss
    def forward(self, x, y):              
        x_vgg, y_vgg = self.vgg(x), self.vgg(y)
        loss = 0
        for i in range(len(x_vgg)):
            # .detach()返回一个新的从当前图中分离的 Variable，返回的 Variable 永远不会需要梯度
            # 可以用于以该变量为输入部分网络求梯度，而不影响y_vgg[]本身
            loss += self.weights[i] * self.criterion(x_vgg[i], y_vgg[i].detach())
        return loss

##############################################################################
# Generator
##############################################################################
# 局部增强器（论文中的G2）
class LocalEnhancer(nn.Module):
    def __init__(self, input_nc, output_nc, ngf=32, n_downsample_global=3, n_blocks_global=9, 
                 n_local_enhancers=1, n_blocks_local=3, norm_layer=nn.BatchNorm2d, padding_type='reflect'):        
        super(LocalEnhancer, self).__init__()
        self.n_local_enhancers = n_local_enhancers
        
        ###### global generator model #####
        # G1 model
        ngf_global = ngf * (2**n_local_enhancers)  # =64
        model_global = GlobalGenerator(input_nc, output_nc, ngf_global, n_downsample_global, n_blocks_global, norm_layer).model        
        model_global = [model_global[i] for i in range(len(model_global)-3)] # get rid of final convolution layers
                                                                             # 最后一层的输出[64,512,512]
        self.model = nn.Sequential(*model_global)                

        ###### local enhancer layers #####
        for n in range(1, n_local_enhancers+1): # =2
            ### downsample            
            ngf_global = ngf * (2**(n_local_enhancers-n))
            model_downsample = [nn.ReflectionPad2d(3), nn.Conv2d(input_nc, ngf_global, kernel_size=7, padding=0), 
                                norm_layer(ngf_global), nn.ReLU(True),
                                nn.Conv2d(ngf_global, ngf_global * 2, kernel_size=3, stride=2, padding=1), 
                                norm_layer(ngf_global * 2), nn.ReLU(True)]
            ### residual blocks
            # model_upsample在此处定义，在 .forward 里使用
            model_upsample = []
            for i in range(n_blocks_local):  # =3
                model_upsample += [ResnetBlock(ngf_global * 2, padding_type=padding_type, norm_layer=norm_layer)]

            ### upsample
            model_upsample += [nn.ConvTranspose2d(ngf_global * 2, ngf_global, kernel_size=3, stride=2, padding=1, output_padding=1), 
                               norm_layer(ngf_global), nn.ReLU(True)]      

            ### final convolution
            if n == n_local_enhancers:                
                model_upsample += [nn.ReflectionPad2d(3), nn.Conv2d(ngf, output_nc, kernel_size=7, padding=0), nn.Tanh()]                       

            # 为中间层命名
            setattr(self, 'model'+str(n)+'_1', nn.Sequential(*model_downsample))
            setattr(self, 'model'+str(n)+'_2', nn.Sequential(*model_upsample))

        # 平均池化，输出 y = (x+2*1-3)/2 + 1，下采样
        self.downsample = nn.AvgPool2d(3, stride=2, padding=[1, 1], count_include_pad=False)

    def forward(self, input): 
        ### create input pyramid
        # （缺省）构建二组不同的输入
        # 通过平均池化，第二组输入尺寸降低1/2
        input_downsampled = [input]
        for i in range(self.n_local_enhancers):  # =1
            input_downsampled.append(self.downsample(input_downsampled[-1])) # [-1]取最后一个元素

        ### output at coarest level
        # 论文中G1输出的特征图
        output_prev = self.model(input_downsampled[-1])        
        ### build up one layer at a time
        # coarse to fine，G1输出的特征图与G2(F)输出的特征图求和，作为model_upsample()的输入
        # G2（F）缺省为只有一层，即：n_local_enhancers=1
        for n_local_enhancers in range(1, self.n_local_enhancers+1):  # =2
            # 取出各中间层
            model_downsample = getattr(self, 'model'+str(n_local_enhancers)+'_1')
            model_upsample = getattr(self, 'model'+str(n_local_enhancers)+'_2')
            # 确定输入
            input_i = input_downsampled[self.n_local_enhancers-n_local_enhancers]  # 1-1 = 0
            # 生成输出
            output_prev = model_upsample(model_downsample(input_i) + output_prev)
        return output_prev

# 全局生成器（论文中的G1）
class GlobalGenerator(nn.Module):
    def __init__(self, input_nc, output_nc, ngf=64, n_downsampling=3, n_blocks=9, norm_layer=nn.BatchNorm2d, 
                 padding_type='reflect'):
        assert(n_blocks >= 0)
        super(GlobalGenerator, self).__init__()        
        activation = nn.ReLU(True)        

        # 第一层，用的是zero_padding
        # 因为第一层用的是7x7的卷积核、padding=0，而 512%7 = 1，因此需要补充6个像素，镜像填充ReflectionPad2d(3)
        # [3,512,512]->[64,512,512]，ngf=64
        model = [nn.ReflectionPad2d(3), nn.Conv2d(input_nc, ngf, kernel_size=7, padding=0), norm_layer(ngf), activation]

        # 下采样，每一层卷积的stride都是2，n_downsampling=3
        ### downsample，stride=2
        # [64,512,512]->[512,64,64]
        for i in range(n_downsampling):
            mult = 2**i
            model += [nn.Conv2d(ngf * mult, ngf * mult * 2, kernel_size=3, stride=2, padding=1),
                      norm_layer(ngf * mult * 2), activation]

        # 残差块，残差块不改变分辨率
        ### resnet blocks
        # dim=512
        mult = 2**n_downsampling
        for i in range(n_blocks):
            model += [ResnetBlock(ngf * mult, padding_type=padding_type, activation=activation, norm_layer=norm_layer)]

        # 和下采样数目一样的上采样部分，上采样部分不像Unet结构，没有用到下采样得到的特征图
        ### upsample，使用转置卷积函数ConvTranspoese2d()，stride=2
        for i in range(n_downsampling):
            mult = 2**(n_downsampling - i)
            model += [nn.ConvTranspose2d(ngf * mult, int(ngf * mult / 2), kernel_size=3, stride=2, padding=1, output_padding=1),
                       norm_layer(int(ngf * mult / 2)), activation]

        # 模型的输出层。这里没有使用归一化
        # [64,512,512]->[3,512,512]
        model += [nn.ReflectionPad2d(3), nn.Conv2d(ngf, output_nc, kernel_size=7, padding=0), nn.Tanh()]        
        self.model = nn.Sequential(*model)
            
    def forward(self, input):
        return self.model(input)             
        
# Define a resnet block
# 定义残差块
class ResnetBlock(nn.Module):
    def __init__(self, dim, padding_type, norm_layer, activation=nn.ReLU(True), use_dropout=False):
        super(ResnetBlock, self).__init__()
        self.conv_block = self.build_conv_block(dim, padding_type, norm_layer, activation, use_dropout)

    def build_conv_block(self, dim, padding_type, norm_layer, activation, use_dropout):
        conv_block = []
        p = 0
        if padding_type == 'reflect':
            conv_block += [nn.ReflectionPad2d(1)]
        elif padding_type == 'replicate':
            conv_block += [nn.ReplicationPad2d(1)]
        elif padding_type == 'zero':
            p = 1
        else:
            raise NotImplementedError('padding [%s] is not implemented' % padding_type)

        conv_block += [nn.Conv2d(dim, dim, kernel_size=3, padding=p),
                       norm_layer(dim),
                       activation]
        if use_dropout:
            conv_block += [nn.Dropout(0.5)]

        p = 0
        if padding_type == 'reflect':
            conv_block += [nn.ReflectionPad2d(1)]
        elif padding_type == 'replicate':
            conv_block += [nn.ReplicationPad2d(1)]
        elif padding_type == 'zero':
            p = 1
        else:
            raise NotImplementedError('padding [%s] is not implemented' % padding_type)
        conv_block += [nn.Conv2d(dim, dim, kernel_size=3, padding=p),
                       norm_layer(dim)]

        return nn.Sequential(*conv_block)

    def forward(self, x):
        out = x + self.conv_block(x)
        return out

# 编码器网络E，生成低维特征，作为生成器网络的输入
# 这是一个标准的编解码器网络，添加了一个实例级平均池层，以计算对象实例的平均特性（找到每一类对象的多个实例）
class Encoder(nn.Module):
    def __init__(self, input_nc, output_nc, ngf=32, n_downsampling=4, norm_layer=nn.BatchNorm2d):
        super(Encoder, self).__init__()        
        self.output_nc = output_nc        

        model = [nn.ReflectionPad2d(3), nn.Conv2d(input_nc, ngf, kernel_size=7, padding=0), 
                 norm_layer(ngf), nn.ReLU(True)]             
        ### downsample，stride=2
        for i in range(n_downsampling):
            mult = 2**i
            model += [nn.Conv2d(ngf * mult, ngf * mult * 2, kernel_size=3, stride=2, padding=1),
                      norm_layer(ngf * mult * 2), nn.ReLU(True)]

        ### upsample，使用转置卷积函数ConvTranspose2d()，stride=2
        for i in range(n_downsampling):
            mult = 2**(n_downsampling - i)
            model += [nn.ConvTranspose2d(ngf * mult, int(ngf * mult / 2), kernel_size=3, stride=2, padding=1, output_padding=1),
                       norm_layer(int(ngf * mult / 2)), nn.ReLU(True)]        

        model += [nn.ReflectionPad2d(3), nn.Conv2d(ngf, output_nc, kernel_size=7, padding=0), nn.Tanh()]
        self.model = nn.Sequential(*model) 

    def forward(self, input, inst):
        outputs = self.model(input)

        # instance-wise average pooling
        outputs_mean = outputs.clone()
        inst_list = np.unique(inst.cpu().numpy().astype(int))  # instance list
        for i in inst_list:
            for b in range(input.size()[0]):  # 对HW做平均池化
                indices = (inst[b:b+1] == int(i)).nonzero() # n x 4，nonzero()返回的是数组中非零元素的位置
                for j in range(self.output_nc): # 每个feature map单独计算
                    output_ins = outputs[indices[:,0] + b, indices[:,1] + j, indices[:,2], indices[:,3]]                    
                    mean_feat = torch.mean(output_ins).expand_as(output_ins)                                        
                    outputs_mean[indices[:,0] + b, indices[:,1] + j, indices[:,2], indices[:,3]] = mean_feat                       
        return outputs_mean

#  多尺度判别器，基于鉴别器的特征匹配损失函数，用来改善GAN损失函数（提高稳定型和优化效率）
class MultiscaleDiscriminator(nn.Module):
    def __init__(self, input_nc, ndf=64, n_layers=3, norm_layer=nn.BatchNorm2d, 
                 use_sigmoid=False, num_D=3, getIntermFeat=False):
        super(MultiscaleDiscriminator, self).__init__()
        self.num_D = num_D
        self.n_layers = n_layers
        self.getIntermFeat = getIntermFeat

        # 生成的NLayerDiscriminator类，被设置（恰当地说，是“命名”）为当前类(self)的一个属性
        # 生成num_D个NLayerDiscriminator
        for i in range(num_D):
            netD = NLayerDiscriminator(input_nc, ndf, n_layers, norm_layer, use_sigmoid, getIntermFeat)
            if getIntermFeat:                                
                for j in range(n_layers+2):
                    # setattr() 函数对应函数 getattr()，用于设置属性值
                    setattr(self, 'scale'+str(i)+'_layer'+str(j), getattr(netD, 'model'+str(j)))                                   
            else:
                setattr(self, 'layer'+str(i), netD.model)

        self.downsample = nn.AvgPool2d(3, stride=2, padding=[1, 1], count_include_pad=False)  # 平均池化，下采样

    def singleD_forward(self, model, input):
        if self.getIntermFeat:
            result = [input]
            for i in range(len(model)):
                result.append(model[i](result[-1]))
            return result[1:]
        else:
            return [model(input)]

    # D的前向传播
    def forward(self, input):        
        num_D = self.num_D
        result = []
        input_downsampled = input
        # 逐一下采样，生成多个不同尺度的输入，并经singleD_forward()生成不同尺度的输出
        for i in range(num_D):
            if self.getIntermFeat:
                model = [getattr(self, 'scale'+str(num_D-1-i)+'_layer'+str(j)) for j in range(self.n_layers+2)]
            else:
                model = getattr(self, 'layer'+str(num_D-1-i))
            result.append(self.singleD_forward(model, input_downsampled))
            if i != (num_D-1):
                input_downsampled = self.downsample(input_downsampled)
        return result

# 用指定的参数定义PatchGAN鉴别器（只定义网络，loss函数在class Pix2PixHDModel()中定义）
# Defines the PatchGAN discriminator with the specified arguments.
class NLayerDiscriminator(nn.Module):
    def __init__(self, input_nc, ndf=64, n_layers=3, norm_layer=nn.BatchNorm2d, use_sigmoid=False, getIntermFeat=False):
        super(NLayerDiscriminator, self).__init__()
        self.getIntermFeat = getIntermFeat
        self.n_layers = n_layers

        kw = 4
        padw = int(np.ceil((kw-1.0)/2))  # =2，np.ceil()计算大于等于该值的最小整数
        # [3, 512, 512] -> [64, 257, 257]
        sequence = [[nn.Conv2d(input_nc, ndf, kernel_size=kw, stride=2, padding=padw), nn.LeakyReLU(0.2, True)]]

        nf = ndf
        for n in range(1, n_layers):
            nf_prev = nf
            nf = min(nf * 2, 512)
            sequence += [[
                nn.Conv2d(nf_prev, nf, kernel_size=kw, stride=2, padding=padw),
                norm_layer(nf), nn.LeakyReLU(0.2, True)
            ]]

        nf_prev = nf
        nf = min(nf * 2, 512)
        sequence += [[
            nn.Conv2d(nf_prev, nf, kernel_size=kw, stride=1, padding=padw),
            norm_layer(nf),
            nn.LeakyReLU(0.2, True)
        ]]

        sequence += [[nn.Conv2d(nf, 1, kernel_size=kw, stride=1, padding=padw)]]

        if use_sigmoid:
            sequence += [[nn.Sigmoid()]]

        # 命名，以方便取出每一个中间层（计算feature mapping loss会用到）
        if getIntermFeat:
            for n in range(len(sequence)):
                setattr(self, 'model'+str(n), nn.Sequential(*sequence[n]))
        else:
            sequence_stream = []
            for n in range(len(sequence)):
                sequence_stream += sequence[n]
            self.model = nn.Sequential(*sequence_stream)

    def forward(self, input):
        if self.getIntermFeat:
            res = [input]
            for n in range(self.n_layers+2):
                model = getattr(self, 'model'+str(n))
                res.append(model(res[-1]))
            return res[1:]
        else:
            return self.model(input)        

from torchvision import models
# VGG19，定义模型的5个切片（只用到0--29层）
class Vgg19(torch.nn.Module):
    def __init__(self, requires_grad=False):
        super(Vgg19, self).__init__()
        vgg_pretrained_features = models.vgg19(pretrained=True).features
        self.slice1 = torch.nn.Sequential()
        self.slice2 = torch.nn.Sequential()
        self.slice3 = torch.nn.Sequential()
        self.slice4 = torch.nn.Sequential()
        self.slice5 = torch.nn.Sequential()
        for x in range(2):
            self.slice1.add_module(str(x), vgg_pretrained_features[x])
        for x in range(2, 7):
            self.slice2.add_module(str(x), vgg_pretrained_features[x])
        for x in range(7, 12):
            self.slice3.add_module(str(x), vgg_pretrained_features[x])
        for x in range(12, 21):
            self.slice4.add_module(str(x), vgg_pretrained_features[x])
        for x in range(21, 30):
            self.slice5.add_module(str(x), vgg_pretrained_features[x])
        if not requires_grad:
            for param in self.parameters():
                param.requires_grad = False

    def forward(self, X):
        h_relu1 = self.slice1(X)
        h_relu2 = self.slice2(h_relu1)        
        h_relu3 = self.slice3(h_relu2)        
        h_relu4 = self.slice4(h_relu3)        
        h_relu5 = self.slice5(h_relu4)                
        out = [h_relu1, h_relu2, h_relu3, h_relu4, h_relu5]
        return out

（完）