search

chojw / genb

Generative Bias for Robust Visual Question Answering ( CVPR 2023 )
MIT License
25 stars 1 forks source link

The pattern collapse during training #8

Closed Larry2000error closed 5 months ago

Larry2000error commented 7 months ago

Hi, I am reproducing this work of yours, I had to make some modifications to your code to get it to run due to different hardware drivers. The changes are as follows:

  1. I modified the position of optim_G.step() to move it before optim_D.step().

but I found that I could not reproduce your results. I have examined the code and found that the problem lies in the training of the GAN, I found that following your code when running on my hardware a pattern collapse occurs, i.e. the generator's loss starts close to 0.However discriminator losses are very high. Here the gen_loss and dsc_loss I have recorded using the code shown below: f3c2321d71a7b80c12064f1b4aad120 44356cab6c5f5d0ccdc32dac138abc4

8f727a1398c920d8ac85225df8adc38

I would like to ask you if you have encountered this problem in your implementation? If not, can you please spot where my problem occurs?

chojw commented 5 months ago

Hi, it seems there are some issues that are coming from version differences. and another issue seems to have been discussing this. Please refer to issue 6.