search

NVlabs / stylegan3

Official PyTorch implementation of StyleGAN3
Other
6.3k stars 1.11k forks source link

FID not improving on custom data #226

Open mukundkhanna123 opened 1 year ago

mukundkhanna123 commented 1 year ago

Hello,

I am trying to train StyleGAN3 on 512x512 black and white PNG images. I have a dataset of about 100k and have used the following command to run training. python train.py --cfg=stylegan3-t --data=dataset.zip --gpus=1 --batch 32 --gamma 16 --aug=noaug --cond True

I have prepared the dataset using dataset_tool.py. I have also experimented with a gamma value of 8 and a smaller discriminator learning rate but I am running into two problems. Firstly, I am training on an A40 gpu but the training is very slow with a speed of 146.14 sec/kimg and secondly my FID score stays in the range of 300-500 and shows no sign of convergence.

If you could give me any advice on how I can speed up training and improve my FID score I'd be highly grateful. Thanks in advance!

dloss gloss fid50k

SamSilmarilData commented 7 months ago

Hi, Any luck?

jasuriy commented 2 months ago

@mukundkhanna123 hi did you fix the issue? were you able to train the model with your own dataset?