Speedup on Image to Image problems?

[swamiviv]

I am trying to use ffcv for an image to image prediction problem, Depth Prediction. I am unable to achieve any speed-up in my basic implementation compared to a a naive dataloading mechanism (reading files from disk and iterating over them). What are ways that I can use to understand if ffcv can actually be used to speed up my use case? What are some knobs I can tune to maybe increase the speed of ffcv? num_workers etc

[andrewilyas]

Hi @swamiviv ! You might be interested in the "Bottleneck Doctor" page in the FFCV docs: https://docs.ffcv.io/bottleneck_doctor.html . Hopefully this helps! Another place to look would be the tuning guide, here: https://docs.ffcv.io/parameter_tuning.html