您好，HiFaceGAN找不到GANFeatLoss（I cannot use GANFeatLoss when trying to run HiFaceGAN by BasicSR）

[Lihua20001020]

我在使用您的BasicSR框架运行其中的HiFaceGAN人脸超分辨率模型时，训练前的构建模型阶段报错“KeyError: "No object named 'GANFeatLoss' found in 'loss' registry!"“，请问这个GANFeatLoss在哪个文件中，是否已经实现了呢？

When I was using your BasicSR framework to run the HiFaceGAN face super-resolution model, I encountered an error during the model building stage before training: "KeyError: "No object named 'GANFeatLoss' found in 'loss' registry!" Could you please tell me in which file the GANFeatLoss is located, and if it has already been implemented?

[Lihua20001020]

看起来这几天没人关注，我自己根据HiFaceGAN的论文实现了一下GANFeatLoss。
It seems that no one has paid attention to this question in the past few days, so I implemented the GANFeatLoss based on the HiFaceGAN paper by myself.

import torch
from basicsr.utils.registry import LOSS_REGISTRY
from basicsr.losses.gan_loss import GANLoss

@LOSS_REGISTRY.register()
class GANFeatLoss(GANLoss):

def __init__(self, gan_type, real_label_val=1.0, fake_label_val=0.0, loss_weight=10.0):
    super(GANFeatLoss, self).__init__(gan_type, real_label_val, fake_label_val, loss_weight)
    self.FloatTensor = torch.cuda.FloatTensor
    self.criterionFeat = torch.nn.L1Loss()
    self.loss_weight = loss_weight

def forward(self, pred_fake, pred_real):

    num_D = len(pred_fake)                                                                                                                                    
    GAN_Feat_loss = 0

    for i in range(num_D):  # for each discriminator                                                                
        num_intermediate_outputs = len(pred_fake[i]) - 1                                                            
        for j in range(num_intermediate_outputs):  # for each layer output                                          
            unweighted_loss = self.criterionFeat(pred_fake[i][j], pred_real[i][j].detach())                         
            GAN_Feat_loss += unweighted_loss * self.loss_weight / num_D                                             

    return GAN_Feat_loss