DistSup

A framework for unsupervised and distant-supervised representation learning with variational autoencoders (VQ-VAE, SOM-VAE, etc), brought to life during the 2019 Sixth Frederick Jelinek Memorial Summer Workshop.

Papers:

• Unsupervised Neural Segmentation and Clustering for Unit Discovery in Sequential Data (see egs/segmental)
• Robust Training of Vector Quantized Bottleneck Models (under review; see egs/robustvq)
• Neural Variational representation learning for spoken language (under review; TBA)

Docker

The easiest way to begin training is to build a Docker container

docker build --tag distsup:latest .
docker run distsup:latest

Installation

We supply all dependencies in a conda environment. Read how to set up the environment.

Training

To get started, train a simple model on MNIST dataset. See egs/ for more examples.

Make sure to load the environment. You can train models using the train.py/train.sh script:

./train.sh egs/mnist/yamls/mlp.yaml runs/mnist/mlp

train.sh is a tiny wrapper around train.py which saves the source code and captures all output to a file, helping to recover the settings of finished experiments.

Some useful command line options are:

• -c LAST resumes training from the last checkpoint. (It is safe to always use it, even during the first training run).
• -m param val overrides parameters from an experiment .yaml file; for instance, to disable weight noise:

./train.sh -c LAST egs/mnist/yamls/mlp.yaml runs/mnist -m Trainer.weight_noise 0.0

For training ScribbleLens models, download the data with

bash egs/scribblelens/download_data.sh

Evaluating models

A saved checkpoint can be loaded and its evaluation metrics run with:

python evaluate.py runs/mnist/mlp/

Visualizing training

Progress of training is logged to Tensorboard. To view training stats run tensorboard --logdir PATH_TO_RUN_FOLDER.

Contributing

• All contributions are welcomed!
• Neural modules lay out the data as NWHC (that is batch_size x width x heigth x channel or equivalently batch_size x time x frequency x channel) with setting width/time being the variable dimension and setting the H dimension to 1 for 1D modules.
• Please use distsup.utils.safe_squeeze to remove it for an adidtional protection (torch.squeeze silently doesn't squeeze in such case).