search

eladrich / pixel2style2pixel

Official Implementation for "Encoding in Style: a StyleGAN Encoder for Image-to-Image Translation" (CVPR 2021) presenting the pixel2style2pixel (pSp) framework
https://eladrich.github.io/pixel2style2pixel/
MIT License
3.2k stars 568 forks source link

Can large batchsize accelerate convergence? #292

Closed caopulan closed 2 years ago

caopulan commented 2 years ago

I tried large batch-size (i.e., 16 and 64) with linearly increasing learning rate lr=1e-4*(bs/8). However, it can not accelerate the convergence, both happening in pSp and e4e. Have you ever tested performance in large batchsize or can you give me some suggestions?

yuval-alaluf commented 2 years ago

We did not test performance with larger batch sizes. If you do change the batch size, you should adjust the loss lambdas accordingly. This could explain why you are not seeing accelerated convergence.