torchvggish-gpu

Re-Implementation of Google Research's VGGish model used for extracting audio features using Pytorch with GPU support.

A re-implementation of VGGish^[1], a feature embedding frontend for audio classification models, using Pytorch with GPU support. This code is fully based on torchvggish^[2].

Usage

import torch
from torchvggish_gpu import vggish
import vggish_input

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# For GPU support, the device must be cuda

embedding_model = vggish()
embedding_model.to(device)
embedding_model.eval()
example = vggish_input.wavfile_to_examples("bus_chatter.wav")
example = example.to(device)
audio_embeddings = embedding_model.forward(example)

---

[1] S. Hershey et al., ‘CNN Architectures for Large-Scale Audio Classification’, in International Conference on Acoustics, Speech and Signal Processing (ICASSP),2017. Available: https://arxiv.org/abs/1609.09430, https://ai.google/research/pubs/pub45611

[2] Harri Taylor et al., ‘Pytorch port of Google Research's VGGish model used for extracting audio features’, v0.1, Sep 27, 2019. Available: https://github.com/harritaylor/torchvggish/releases/tag/v0.1