ZeroDiffusion

ZeroDiffusion — a zero-shot learning diffusion method for audio classification and generation experiments.

This repository contains training and evaluation scripts that rely on precomputed audio embeddings. The workflow is:

• Train or obtain an audio embedding model (see data/audio_embeddings/README.md).
• Precompute embeddings and save them as a pickle in the data/ folder (see data/README.md).
• Run the ZeroDiffusion training scripts.

Prerequisites

• Python 3.8+ (use a virtual environment)
• Install dependencies:

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Quick start

1. Prepare audio embeddings and place the dataset pickle under data/.
2. Train the model (example):

python train.py ../data/ESC_50_synonyms/fold04.pickle --dataset ESC-50 --split fold0

There are additional training/evaluation scripts in the repo:

• train.py — primary ZeroDiffusion training entrypoint.
• train_ddpm.py, train_denoise.py — DDPM / denoising variants.
• run.sh, run_ddpm.sh — convenience wrappers for common runs.

Repository layout

• data/ — dataset helpers, embedding pickles, and dataset README.
• dataloader/ — data loading and diffusion-related dataloaders.
• models/ — model implementations (classifier, ddpm, denoise).
• results/ — generated logs and outputs from experiments.

Notes

• See data/README.md and audio_embeddings/README.md for dataset and embedding preparation details.
• Adjust hyperparameters in the training scripts or pass flags on the command line.

If you'd like, I can also:

• Add example commands for the other scripts.
• Add a minimal config file for common hyperparameters.
• Create a short tutorial notebook demonstrating end-to-end runs.