search

omegahh / DeepHiC

A GAN-based method for Enhancing Hi-C data
MIT License
28 stars 8 forks source link

bad results observed in training and evaluating DeepHiC at 5kb #9

Open zhyanlin opened 1 year ago

zhyanlin commented 1 year ago

Hello,

I'm trying to train a DeepHiC model at 5kb resolution.

I used 1/16 downsampled GM12878 data as input and trained DeepHiC with default parameters. During training, I observed both loss_D and loss_G were not improved for a very long time. [400/1000] Loss_D: 0.0000 Loss_G: 0.0013 D(x): 1.0000 D(G(z)): 0.0000 I trained it for a very long time. It reached 0.0000, and 0.0013 around 100 epochs. The Best SSIM is ~0.901. Here is an example of the prediction of a 180x180 window. Could you suggest how to update the parameters to train a better model or share a trained model at 5kb if you have one? It would be helpful if you could also share training plots such as loss v.s. epoch so I can compare it with mine.

image

Thanks!

YuantingMaSC commented 1 month ago

It looks like your reproduction of the paper went much smoother than mine, I'd like to ask you how you drew your SSMI and the example HiC plots, I didn't find the corresponding scripts in the project