search

microsoft / StyleSwin

[CVPR 2022] StyleSwin: Transformer-based GAN for High-resolution Image Generation
https://arxiv.org/abs/2112.10762
MIT License
503 stars 49 forks source link

About Automatic Mixed Precision #12

Closed HAN-oQo closed 2 years ago

HAN-oQo commented 2 years ago

Thank you for awesome research and code release! Is there any reason that you don't use automatic mixed precision package of pytorch? Did it lower the performance of model when you use it?

ForeverFancy commented 2 years ago

Hi, we have tried to use amp in the early experiments. However, the training is not stable, and the speedup is not significant (the cuda ops need fp32 precision). Therefore, we did not use it for the final version.

HAN-oQo commented 2 years ago

Thank you!