Hi,
I have a model that in itself takes relatively little GPU memory (around 1Gb, 1/8th of what I have available), however the dataset is comparatively large (around 40GB, which fits in RAM). Since loading the data takes a long time, I load data lazily and keep everything in RAM - meaning the first epoch takes around 30 minutes to run, but every epoch after that takes less than one minute.
I’m tuning hyperparameters and thus need to train many iterations of the model. My GPU could easily train 7/8 models in parallel, however since I only have 96GB of ram, I can only fit the dataset twice in memory, meaning I can only train two models at a time.
Would it be possible/how can I use a single Dataset/Dataloader for multiple models so that I don’t need to have duplicate data in memory?
Thanks!