@SamuelLarkin suggests fuse-mounting a zip archive or a new dataloader that preprocesses the audio on-the-fly (if it doesn't slowdown training aka makes the trainer wait on preprocessing) to help circumvent the perennial inode problem for large datasets.
@SamuelLarkin suggests fuse-mounting a zip archive or a new dataloader that preprocesses the audio on-the-fly (if it doesn't slowdown training aka makes the trainer wait on preprocessing) to help circumvent the perennial inode problem for large datasets.