search

akanimax / BMSG-GAN

[MSG-GAN] Any body can GAN! Highly stable and robust architecture. Requires little to no hyperparameter tuning. Pytorch Implementation
MIT License
630 stars 105 forks source link

Why smaller batch size results better results? With the epochs increased, the G loss went up? #39

Open AAAeray opened 4 years ago

AAAeray commented 4 years ago

There are 8k+ 256*256 images in my datasets, I set batchsize=32 and trained with 4 12GB GPUs, but it's not as good as I set batchsize = 4 and trained with one GPU,neither the training speed nor the image quality. Why a smaller batch size results better results? Is there a best batchsize? Beside,I set batchsize=4, when epoch>70, g-loss went up obviously,and qualities of generated images went worse.Why did this happen? (Training process was interrupted when epoch = 54, I reloaded weight files and optimer states from epoch 53 ) Thanks!

akanimax commented 4 years ago

Well, there are some known issues regarding the multi-gpu training since the last couple of updates of PyTorch. I will have to investigate this, but it's quite difficult given my schedule right now. The training instability in this resumed training could be because of improper loading of state/weights.