Real readme coming soon. 2020/05/26 snapshot of code.
For now, find the Bayesian VAE and WaveNet reimplementations (in PyTorch) under src/models. Each model has a python script in src/ for running training instances with that model. For example
python run_vae.py --log_interval batch
from the src folder trains a newly initialized VAE model on default arguments (which can be viewed using the --help or -h flag), with logging information per batch update (basically a training progress bar). We currently get the best scores on VAE training 5 models, each with sparse interactions (dict flag) and randomly weighted samples.
python run_vae.py -dict -rws
Then, use the evaluate_ensemble.py script in src to make ensemble predictions over the trained models. This script takes one or more directories as argument and looks for model files (all files with .torch suffix) to include in the ensemble.
Place aligned data files under src/data/files/alignments/. The alignment data used by default is from the DeepSequence repository, which can easily be downloaded and extracted using their script, as provided here as well. Navigate to src/data/files/ and run this script to download and extract alignment files:
curl -o alignments.tar.gz https://marks.hms.harvard.edu/deepsequence/alignments.tar.gz
tar -xvf alignments.tar.gz
rm alignments.tar.gz