VGGVox-PyTorch

Implementing VGGVox for VoxCeleb1 dataset in PyTorch.

Train

pip install -r requirements.txt
python3 train.py --dir ./data/

Specify data dir with --dir

Notes

• 81.79% Top-1 & 93.17 Top-5 Test-set accuracy, pretty satisfactory. Find details in results.txt.
• Training on the V100 takes 4 mins per epoch.

Model

• Run python3 vggm.py for model architecture.
• Best model weights uploaded VGGM300_BEST_140_81.99.pth

What i've done so far:

• [x] All the data preprocessed exactly as author's matlab code. Checked and verified online on matlab
• [x] Random 3s cropped segments for training.
• [x] Copy all hyperparameter... LR, optimizer params, batch size from the author's net.
• [x] Stabilize PyTorch's BatchNorm and test variants. Improved results by a small percentage.
• [x] Try onesided spectrogram input as mentioned on the author's github.
• [ ] Port the authors network from matlab and train. The matlab model has 1300 outputs dimension, will test it later.
• [ ] Copy weights from the matlab network and test.

References and Citations:

• VGGVox
• linhdvu14's vggvox-speaker-identification

• jameslyons's python_speech_features

  
  @InProceedings{Nagrani17,
  author       = "Nagrani, A. and Chung, J.~S. and Zisserman, A.",
  title        = "VoxCeleb: a large-scale speaker identification dataset",
  booktitle    = "INTERSPEECH",
  year         = "2017",
  }

@InProceedings{Nagrani17, author = "Chung, J.~S. and Nagrani, A. and Zisserman, A.", title = "VoxCeleb2: Deep Speaker Recognition", booktitle = "INTERSPEECH", year = "2018", }