ABnet3

Representation learning package using side information, system for subword modeling for Zeroresource challenge.

Overview

Build Representation for speech frames based on side information. Composed of different modules :

• model.py
• loss.py
• sampler.py
• trainer.py
• embedder.py
• utils.py
• features.py

Installation of the package

Using conda

To install the ABnet3 package, you can use Anaconda, and either create a conda environment:

conda env create --name abnet3 python=3.6 -f environment.yml

or use a conda environment you already have with python 3 : conda env update -f environment.yml

To install with GPU support (replace cuda75 with your version of cuda)

conda install  pytorch=0.2 cuda75 -c pytorch

Using pip

• install the version 0.2.0 of pytorch for your hardware (http://pytorch.org/previous-versions/)

• install the pip packages : pip install -r requirements.txt Once all the necessary packages are installed, simply launch:

Run abnet3 installation

python setup.py build && python setup.py install

If you want to work on ABnet3 and develop your own modules, instead of:

python setup.py install

you can launch:

python setup.py develop

Tensorboard vizualisation

The package tensorboardX needs to be installed to train the model: pip install tensorboardX.

The package will save train / dev loss during training. To vizualise them :

• Install tensorboard (conda install tensorflow tensorflow-tensorboard)

• run tensorboard --logdir path/to/logdir. The default logdir is ./run in the current directory.

Documentation

You can see examples for running the gridsearch and replicating our results in the repository https://github.com/Rachine/sampling_siamese2018

The cli documentation is here https://coml.lscp.ens.fr/git/Rachine/abnet3/src/master/gridsearch.md

Tests

The package comes with a unit-tests suit. To run it, first install pytest on your Python environment:

pip install pytest
pytest test/

References

.. [1] Riad, R., Dancette, C., Karadayi, J., Zeghidour, N., Schatz, T., Dupoux, E.
       *Sampling strategies in Siamese Networks for unsupervised speech representation learning.*
       In Nineteenth Annual Conference of the International Speech Communication Association

.. [2] Thiolliere, R., Dunbar, E., Synnaeve, G., Versteegh, M., & Dupoux, E.
       *A hybrid dynamic time warping-deep neural network architecture for unsupervised acoustic modeling.*
       In Sixteenth Annual Conference of the International Speech Communication Association

.. [3] Zeghidour, N., Synnaeve, G., Usunier, N. & Dupoux, E.
       *Joint Learning of Speaker and Phonetic Similarities with Siamese Networks.*
       In: INTERSPEECH-2016, (pp 1295-1299)

Acknowledgments

A part of the code is inspired from the previous version in Theano of ABnet, and the examples in Pytorch