CLAM Model Training Guide

This README provides comprehensive instructions for setting up and training the CLAM model.

1. Dataset Acquisition and Organization

The first step is to gather and organize your audio data.

• Download the Dataset: All shareable data and links to publicly downloadable content for this project can be found on Hugging Face: https://huggingface.co/datasets/anonymous2212/MoM-CLAM-dataset
• Organize Audio Files: After downloading, ensure all audio files are placed into their corresponding folders. The csv metadata file provides the necessary mapping for this organization.

2. Feature Embedding Extraction

Once your dataset is prepared, you'll need to extract features using the provided scripts.

• Run Extraction Scripts:
  • Execute extractors/mertextraction.py
  • Execute extractors/wave2vec2extract.py
• Save Embeddings:
  • For AI-generated music, save the MERT embeddings to ai_generated_music_mert/ and the Wav2Vec2 embeddings to ai_generated_music_wav2vec2/.
  • Apply the same saving conventions for "real" music, creating similar directories (e.g., real_music_mert/).
• Important: Refer to the specific instructions within the extractors/ files, (arguments) for detailed usage and configuration of these scripts.

3. Model Training

With features extracted, you're ready to train the CLAM model.

• Initiate Training: Run one of the following files to begin the training process and obtain the final results:
  • train_triplet_loss.ipynb (Jupyter Notebook for interactive training)
  • triplet_train.py (Python script for command-line execution)

Instructions for running CLAM

• Organise and download all the audio files in the respective folders as shown in the csv metadeta file.
• Dataset - https://huggingface.co/datasets/anonymous2212/MoM-CLAM-dataset, it contains all the shareable data along with links for other downloadable public content.
• run "extractors/mertextraction.py" and "extractors/wave2vec2extract.py" and save embeddings in - "ai_generated_music_mert", "ai_generated_music_wav2vec2" and similar for real music using the instructions provided in the extractors file.
• then run train_triplet_loss.ipynb or triplet_train.py to achieve final results.