remove transfers to cpu in dataloader

balthazarneveu / blind-deblurring-from-synthetic-data

MVA ENS Paris Saclay - Image restoration project on deblurring learnt on deadleaves

3 stars 0 forks source link

Closed balthazarneveu closed 5 months ago

balthazarneveu commented 6 months ago

In current gpu dataloader, we end up going back and forth between GPU and CPU .

Avoid this at least for AWGN task and add noise in torch with random tensors instead of CPU numpy
Factor degradation code between dataloaders and inference blocks (interactive ?) to avoid inconsistencies

balthazarneveu commented 5 months ago

Done per PR #44 aka the "Degradation module" 🙏 thanks @Jamy-L