search

JustinhoCHN / SRGAN_Wasserstein

Apply Waseerstein GAN into SRGAN, a deep learning super resolution model
420 stars 102 forks source link

w_loss won't go down #16

Open lf-openthos opened 5 years ago

lf-openthos commented 5 years ago

Did everything just like readme.

Epoch [854/1500] 27 time: 1.3290s, W_loss: -6.52173805 g_loss: 0.03743032 (mse: 0.021245 vgg: 0.012784 adv: 0.003402) Epoch [854/1500] 28 time: 1.3205s, W_loss: -6.52044964 g_loss: 0.03023629 (mse: 0.013892 vgg: 0.012944 adv: 0.003401) Epoch [854/1500] 29 time: 1.3281s, W_loss: -6.47881317 g_loss: 0.02639372 (mse: 0.012046 vgg: 0.010970 adv: 0.003378) Epoch [854/1500] 30 time: 1.3177s, W_loss: -6.50874901 g_loss: 0.02752733 (mse: 0.013138 vgg: 0.010993 adv: 0.003397) Epoch [854/1500] 31 time: 1.3302s, W_loss: -6.49015522 g_loss: 0.03011781 (mse: 0.013609 vgg: 0.013120 adv: 0.003389)

ZhiluDing commented 5 years ago

i also have some problems with the w_loss. I don't understand why the w_loss changes from 6.xxx to -6.xxx (why it is bigger than 1.0 and it is can be negative). and when the w_loss is converge. Hope some one can help me

ponykid commented 2 years ago

I notice the w_loss decreases to -6.0, so I am so confused about that why w_loss can be negative?