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

原创

已于 2022-09-05 07:57:18 修改 · 5.2k 阅读

19 ·

CC 4.0 BY-SA版权

文章标签：

#git #windows #pycharm

于 2021-02-19 19:59:10 首次发布

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_righ