Training on large dataset (10 gbs / 20k )

[MISSEY]

I need to train on a large dataset.

There are no errors while training on a small dataset such as 1GB, but while trying to train a larger dataset which is included in annotation, get struck and Attempting to copy inputs of <function pairwise_iou at > to CPU due to CUDA OOM.

If I wish to change the training loop what things I have to change so that the entire dataset should not be loaded.

Thank you
Best