[FeatureRequest] GPU transforms about rising HOT 2 CLOSED

The problem is, that it creates a separate cuda context for each worker process. However I'm not sure if the approach with coroutines is the best one or if we just implement a queueing system so that we only use one process for GPU transforms and multiple processes for loading and CPU transforms

from rising.

IMO we don't want multiprocessing for the GPU. I did some experiments and it uses way to much memory and is not for what GPUs are meant for in the first place.

I think we need to introduce some point in the data loader where the preprocessed data from the CPU is collected and only transform a single batch at the time on the gpu. The workflow would look roughly like this:

data => cpu processing
data => cpu processing => collect => when batch requested: put on target gpu => transform => network
data => cpu processing

I could imagine the "request batch" part could be realised by python coroutines :)

from rising.