Awesome-Music-Generation

This repository is the implementation of music generation model MMGen.

FAISS Experiments

This module optimizes FAISS (Facebook AI Similarity Search) indexes for audio and melody data in the MMGen music generation model.

Scripts

1. find_best_faiss_params.py: Evaluates FAISS index types (Flat, IVF, HNSW) and their parameters.

   • Calculates intersection rates and search times
   • Outputs: find_best_faiss_params.json

2. save_faiss_params.py: Builds and saves FAISS indexes using selected parameters.

   • Constructs HNSW and IVF indexes
   • Outputs:
     • Indexes in ./OverlapRate_Experiments_result/
     • index_info.json with performance metrics

Data

Required files (in ./Multimodal_Alignment_npy/):
example

• musiccaps_melody_362_trimmed.npy
• musiccaps_audio_362_trimmed.npy

Usage

1. Run python find_best_faiss_params.py to evaluate parameters
2. Manually select optimal parameters from find_best_faiss_params.json
3. Create save_faiss_params.json with the selected parameters
4. Run python save_faiss_params.py to build and save indexes