The present dataset implementation uses a context manager to open an h5 file, load it, and close the file with every __getitem__ call, and additionally computes expensive transformations with each __getitem__ call. If possible, initially loading the entire dataset into memory could speed up training.
The present dataset implementation uses a context manager to open an h5 file, load it, and close the file with every
__getitem__call, and additionally computes expensive transformations with each__getitem__call. If possible, initially loading the entire dataset into memory could speed up training.