Learning music audio representations with limited data

Code for “Learning Music Audio Representations with Limited Data”.

Overview

What happens when we train music audio representation models with very limited data?

We train the following models on subsets of the MagnaTagATune music dataset, ranging from 5 to ~8000 minutes.

Name	Architecture	Param.	Emb. Dim.	Input len.	Input feat.	Mel bins	Paradigm
VGGish	CNN	3.7m	512	3.75s	mel spec.	128	Tagging
MusiCNN	CNN	12.0m	200	3.00s	mel spec.	96	Tagging
AST	Transformer	87.0m	768	5.12s	mel spec.	128	Tagging
CLMR	CNN	2.5m	512	2.68s	waveform	-	SSL Contrastive
TMAE	Transformer	7.2m	256	4.85s	mel spec.	96	SSL Masked

We extract representations from each, along with untrained models, and train downstream models on

We show that, in certain cases,

the representations from untrained and minimally-trained models perform comparatively to those from “fully-trained” models
larger downstream models are able to "recover" performance from untrained and minimally-trained representations
the inherent robustness of representations to noise is bad across the board
the performance gap to "hand-crafted" features is still significant in pitch and instrument recognition

pip install -r requirements.txt

MagnaTagATune will be downloaded automatically if it's not already present in data/MTAT. Each model has a training script, which can be run with:

python train.py --model model_name

where model_name is one of musicnn, vggish, ast, clmr, or tmae.

MagnaTagATune, TinySOL, and Beatport will be downloaded automatically if they're not already present in data/.

python extract_features.py --model model_name --task task_name

where model_name is one of musicnn, vggish, ast, clmr, or tmae, and task_name is one of tagging, pitch, or instrument.

python downstream.py --model model_name --task task_name

where model_name is one of musicnn, vggish, ast, clmr, or tmae, and task_name is one of tagging, pitch, or instrument.

See visualization notebooks.