search

ustclby / Unsupervised-Domain-Specific-Deblurring

Implementation of "Unsupervised Domain-Specific Deblurring via Disentangled Representations"
110 stars 28 forks source link

Gradient Clipping #1

Closed unlugi closed 5 years ago

unlugi commented 5 years ago

Dear author,

In the DRIT model you have taken inspiration from, gradient clipping is applied to stabilize gan training. I couldn't see that in your implementation if I am not mistaken. I was wondering if this is intentional.

Thank you.

unlugi commented 5 years ago

Also, I am training with a different dataset. And after a while, the generator and discriminator losses blow up, and I start to receive nan values. What are your recommendations for adapting the training process for a new dataset?

zabir-nabil commented 4 years ago

@unlugi how did you solve the loss issue can you tell?

liuzhuang1024 commented 3 years ago

I also encountered the same problem. No matter how you do it, there are situations where loss is equal to nan. I don’t know if the author has encountered it during the training.