图像梯度差(image gradient difference loss)函数与hinge损失函数， SSIM损失-210219

1.图像梯度差loss(image gradient difference loss)函数

阅读文献看到梯度差损失函数，在github上搜索，可以查到相关代码，做此记录：

def loss_gradient_difference(real_image,generated): # b x c x h x w
    true_x_shifted_right = real_image[:,:,1:,:]# 32 x 3 x 255 x 256
    true_x_shifted_left = real_image[:,:,:-1,:]
    true_x_gradient = torch.abs(true_x_shifted_left - true_x_shifted_right)

    generated_x_shift_right = generated[:,:,1:,:]# 32 x 3 x 255 x 256
    generated_x_shift_left = generated[:,:,:-1,:]
    generated_x_griednt = torch.abs(generated_x_shift_left - generated_x_shift_right)

    difference_x = true_x_gradient - generated_x_griednt

    loss_x_gradient = (torch.sum(difference_x)**2)/2 # tf.nn.l2_loss(true_x_gradient - generated_x_gradient)

    true_y_shifted_right = real_image[:,:,:,1:]
    true_y_shifted_left = real_image[:,:,:,:-1]
    true_y_gradient = torch.abs(true_y_shifted_left - true_y_shifted_right)

    generated_y_shift_right = generated[:,:,:,1:]
    generated_y_shift_left = generated[:,:,:,:-1]
    generated_y_griednt = torch.abs(generated_y_shift_left - generated_y_shift_right)

    difference_y = true_y_gradient - generated_y_griednt
    loss_y_gradient = (torch.sum(difference_y)**2)/2 # tf.nn.l2_loss(true_y_gradient - generated_y_gradient)

    igdl = loss_x_gradient + loss_y_gradient
    return igdl

后续新的损失函数在此继续补充。
参考网址：https://github.com/Zxl19990529/UGAN-pytorch/blob/40d4842c768e9aaea9f57c8c832c531a6a9d85bb/utils.py

2. Hinge loss 是对地球移动距离的一种拓展

Hinge loss 最初是SVM中的概念，其基本思想是让正例和负例之间的距离尽量大，后来在Geometric GAN中，被迁移到GAN:

对于D来说，只有当D(x) < 1 的正向样本，以及D(G(z)) > -1的负样本才会对结果产生影响。也就是说，只有一些没有被合理区分的样本，才会对梯度产生影响。这种方法可以使训练更加稳定。

def __call__(self, outputs, is_real, is_disc=None):
    if self.type == 'hinge':
        if is_disc:
            if is_real:
                outputs = -outputs
            return self.criterion(1 + outputs).mean()
        else:
            return (-outputs).mean()

    else:
        labels = (self.real_label if is_real else self.fake_label).expand_as(outputs)
        loss = self.criterion(outputs, labels)
        return loss

3. 更多loss

https://github.com/tariktemur/RGAN-and-RaGAN-PyTorch/blob/master/utils.py

https://github.com/VainF/pytorch-msssim

total variation loss:
https://github.com/grorge123/project/blob/dc91f2a219/face2music/magenta/models/image_stylization/learning.py