Different results when changing the data loader type.

facebookresearch / tava

Code for the paper "TAVA Template-free Animatable Volumetric Actors".

BSD 3-Clause "New" or "Revised" License

198 stars 18 forks source link

Different results when changing the data loader type. #5

Closed coreqode closed 1 year ago

coreqode commented 1 year ago

Hi again!!

Is there a reason you have used the torch.utils.data.IterableDataset rather than the torch.utils.data.Dataset? When I ported the code to the latter, training was much faster, but the results were not good. While in the case of the former (default), training is slower. Further what's the use case of cache_n_repeathere.

liruilong940607 commented 1 year ago

I believe torch.utils.data.IterableDataset is inherited from torch.utils.data.Dataset so there shouldn’t be a performance discrepancy.

Cache is for reusing the same image for multiple iterations. I was having IO bottlenecks so just cache the image being load and repeatly sample rays in the sample image for multiple iterations. This is not ideal for performance but gave me some speedup during training

liruilong940607 commented 1 year ago

Close for now as it seems the question is resolved.