voxceleb_triplet-loss

A Pytorch implementation of triplet loss on VoxCeleb1

Train

Using softmax pre-training to initialise the network.

triplets selector: semi-hard

Test

Full length utterances are used in testing. So we can fed into network just one example everytime. But if using multi-GPUs(e.g. 6), you can input 6 examples for each iteration.

Result

embedding_dim=256:

-	EER	m-DCF
triplet loss	7.22	0.639

Usage

Create virtual environment:

conda create -n vox_triplet_loss python=3.7
conda activate vox_triplet_loss

Install pytorch: https://pytorch.org/get-started/locally/
Pretrain:
- Change the train_mode to ‘pretrain’ in constants.py
- Set GPU configuration, data_path and save_path in your computer
- Run the train.py
Use triplet loss to train model:
- Change the train_mode to ‘triplet’ in constants.py

Reference

https://github.com/davidsandberg/facenet

https://github.com/adambielski/siamese-triplet