This repository contains the code used in the paper "Target-specific Dataset Pruning for Compression of Audio Tagging Models"
The proposed method combines model compression and domain adaptation to obtain better efficient audio tagging models. It consists of three steps: Data pruning, knowledge distillation and a final fine-tuning.
This method was used for a submission to Task 1 of the DCASE Challenge 2024
Install requirements manually from the requirements.txt and install the package.
Make sure to modify the system path configuration or add your own in target_distillation/conf/system/*.yaml
Refer to audio-data to obtain a webdataset version of AudioSet.
ESC-50 can be downloaded here.
Create logit datasets for both AudioSet and the target datasets using target_distillation/create_ensemble_embeddings.py.
The notebook in domain_classifier allows to train a domain classifier model for a given dataset. For new datasets, a dataset class needs to be added in target_distillation/data.
With the trained domain classifier, AudioSet can be filtered using the create_wds.py script.
Use the ex_distill.py script to distill a model, a fine-tuning is automatically performed after the training is finished.
If there are some missing files or packages, please let me know.
The project was funded by the Federal Ministry of Education and Research (BMBF) under grant no. 01IS22094E WEST-AI.