Pre-encode

Pre-encode images using frozen encoder for faster decoder training

Notes

This script pre-encodes images using a trained encoder checkpoint, saving the embeddings for faster decoder training.

Usage:

python midi_rae/preencode.py encoder_ckpt=checkpoints/best.pt preencode.output_dir=preencoded/

TODO: - May need a simpler Dataset that returns single images (not pairs) + their filenames - Decide on output format: one .pt per image, or chunked/batched files? - Add config entries for encoder_ckpt path and preencode.output_dir

---

preencode

def preencode(
    cfg:DictConfig
):