MangoRAG

Train

!accelerate launch --num_processes 2 --mixed_precision fp16 /kaggle/working/MangoRAG/train.py --audio-encoder-checkpoint openai/whisper-base --text-encoder-checkpoint google-bert/bert-base-multilingual-uncased --model-name whisper-base-bert-base-multilingual-uncased-512 --neptune-tags whisper-base bert-base-multilingual-uncased BS-48 fp16 EMBED-512 LR-1e-05 TAU-0 --audio-encoder-name whisper --text-encoder-name bert --batch-repeats 4 --batch-size 12 --epochs 20 --logits-scale 0 

Eval

!python3 src/cli/eval.py --audio-encoder-name whisper --text-encoder-name bert  --audio-encoder-checkpoint openai/whisper-tiny --text-encoder-checkpoint google-bert/bert-base-uncased

Or

!python3 src/cli/eval.py --clap-run-id MNGCLAP-386