faceswap-GAN之adversarial_loss_loss（对抗loss）

一，faceswap-GAN之adversarial_loss_loss（对抗loss）

二，adversarial_loss，对抗loss，包含生成loss与分辨loss。

def adversarial_loss(netD, real, fake_abgr, distorted, gan_training="mixup_LSGAN", **weights):   
    alpha = Lambda(lambda x: x[:,:,:, :1])(fake_abgr)
    fake_bgr = Lambda(lambda x: x[:,:,:, 1:])(fake_abgr)
    fake = alpha * fake_bgr + (1-alpha) * distorted #mask combine input&output
    #mixup_LSGA
    if gan_training == "mixup_LSGAN":
        dist = Beta(0.2, 0.2)
        lam = dist.sample()
        mixup = lam * concatenate([real, distorted]) + (1 - lam) * concatenate([fake, distorted]) # fake（mask）&real mixup   
        pred_fake = netD(concatenate([fake, distorted])) #Discriminator
        pred_mixup = netD(mixup)#Discriminator
        loss_D = calc_loss(pred_mixup, lam * K.ones_like(pred_mixup), "l2")#one false and zero true for Discriminator
        loss_G = weights['w_D'] * calc_loss(pred_fake, K.ones_like(pred_fake), "l2")#one true and zero false for generator
        mixup2 = lam * concatenate([real, distorted]) + (1 - lam) * concatenate([fake_bgr, distorted]) # fake（no mask）&real mixup 
        pred_fake_bgr = netD(concatenate([fake_bgr, distorted]))
        pred_mixup2 = netD(mixup2)
        loss_D += calc_loss(pred_mixup2, lam * K.ones_like(pred_mixup2), "l2")#one false and zero true for Discriminator
        loss_G += weights['w_D'] * calc_loss(pred_fake_bgr, K.ones_like(pred_fake_bgr), "l2")#one true and zero false for generator
    elif gan_training == "relativistic_avg_LSGAN":
        real_pred = netD(concatenate([real, distorted]))
        fake_pred = netD(concatenate([fake, distorted]))#mask
        loss_D = K.mean(K.square(real_pred - K.ones_like(fake_pred)))/2  #Discriminator real to fake loss
        loss_D += K.mean(K.square(fake_pred - K.zeros_like(fake_pred)))/2  #Discriminator fake to real loss
        loss_G = weights['w_D'] * K.mean(K.square(fake_pred - K.ones_like(fake_pred))) #generator pred to fake loss
        #no mask
        fake_pred2 = netD(concatenate([fake_bgr, distorted])) #no mask
        loss_D += K.mean(K.square(real_pred - K.mean(fake_pred2,axis=0) - K.ones_like(fake_pred2)))/2 #Discriminator real to fake loss
        loss_D += K.mean(K.square(fake_pred2 - K.mean(real_pred,axis=0) - K.zeros_like(fake_pred2)))/2 #Discriminator fake to real loss
        loss_G += weights['w_D'] * K.mean(K.square(real_pred - K.mean(fake_pred2,axis=0) - K.zeros_like(fake_pred2)))/2 
        loss_G += weights['w_D'] * K.mean(K.square(fake_pred2 - K.mean(real_pred,axis=0) - K.ones_like(fake_pred2)))/2
    else:
        raise ValueError("Receive an unknown GAN training method: {gan_training}")
    return loss_D, loss_G

三，LSGAN（最小二乘Gan）
GAN为对抗生成网络，主要用于图像生成，风格转换等，主要有生成网络，识别网络，再加上目标函数loss组成对抗网络。一般对抗网络的差异主要在于目标函数loss的不同。
GAN就是警察抓小偷，是个博弈，警察总是想着如何分辨小偷和非小偷，而小偷总是想着尽可能地伪装成正常人不被发现。这个博弈达到纳什均衡时，小偷的表现就跟正常人一样，而警察也无法判断“小偷”是不是小偷…这个时候的“小偷”就不能算小偷了？
实践中，我们很难去找一个博弈的纳什均衡，一般转而采用梯度算法优化目标函数（损失函数）。
传统GAN中, D网络和G网络都是用简单的交叉熵loss做更新, 最小二乘GAN则用最小二乘(Least Squares) Loss 做更新:

选择最小二乘Loss做更新有两个好处, 1. 更严格地惩罚远离数据集的离群Fake sample, 使得生成图片更接近真实数据(同时图像也更清晰) 2. 最小二乘保证离群sample惩罚更大, 解决了原本GAN训练不充分(不稳定)的问题:

但缺点也是明显的, LSGAN对离离群点的过度惩罚, 可能导致样本生成的”多样性”降低, 生成样本很可能只是对真实样本的简单”模仿”和细微改动.
但LSGAN特别适合换脸，换脸主要还是细微的改变。