Easy interface for extracting features from audio models.

Instructions:

• Git clone this repository and its submodules:

  git clone --recurse-submodules https://github.com/mrpep/easy-audio-embeddings.git
  

• Run install.sh to apply patches to BYOL-A models
• Run pip install -e .

Usage:

from easyaudio.hub import get_model
import numpy as np
import torch

model = get_model('BEATs_iter3')
features = model.extract_activations_from_array(np.random.randn(16000)) #Extract features from numpy array
features = model.extract_activations_from_array(torch.randn((16000,))) #From torch tensor
features = model.extract_activations_from_filename('example.wav') #Given a wav filename

#features is a list of tensors corresponding to the activations from each layer. Each activation has shape (T,D)

Available models:

BEATs:

Paper: https://arxiv.org/abs/2212.09058

Official code: https://github.com/microsoft/unilm/tree/master/beats

Available models: BEATs_iter1, BEATs_iter2, BEATs_iter3, BEATs_iter3+_AS20K, BEATs_iter3+_AS2M

Extraction points: activations are extracted from the transformer input and the output of each transformer layer.

BYOL-A

Paper: https://arxiv.org/abs/2204.07402

Official code: https://github.com/nttcslab/byol-a

Available models: byola_512, byola_1024, byola_2048

Extraction points: activations are extracted from the input of each MaxPooling2D layer and the output of each Linear layer. The last activation in the list corresponds to the original BYOL-A features (final output) that is pooled over time/freq.

EnCodecMAE

Paper: https://arxiv.org/abs/2309.07391

Official code: https://github.com/habla-liaa/encodecmae

Available models: encodecmae_base, encodecmae_base-st, encodecmae_large, encodecmae_large-st, encodecmae_small

Extraction points: activations are extracted from the output of each transformer layer of the Encoder network.