Closed CodeBlo closed 2 years ago
Hi! We trained all the 256x256 models on 4 V100s with 32 GB each for ~2 days. It is very similar in training time to StyleGAN2-ADA. When training on a custom dataset, please keep in mind that the model.generator.time_enc.min_period_len
parameter affects the motion quality a lot: for example, we used model.generator.time_enc.min_period_len=16
for Face Forensics and model.generator.time_enc.min_period_len=256
for SkyTimelapse. In some sense, it reflects the speed of motion — the smaller it is, the more high-frequency the motions would be
Thanks a lot!!!
Welcome! Feel free to ask any further questions if you'll face any issues. Oh, I forgot to add that in fact our model does not need 32GB of memory when trained on 4 GPUs. In fact, it consumes <10GB of memory on each GPU.
It would be really nice to know what are the minimum system requirements to train a custom dataset.
Thanks!