This repository provides an implementation of the algorithm introduced in Modular Universal Reparameterization: Deep Multi-task Learning across Diverse Domains, NeurIPS 2019.
cd muir
mkdir results
pip install -r requirements.txt
export PYTHONPATH=$PYTHONPATH:~/muir/pytorch/
The code assumes datasets are downloaded and placed in ~/hyperdatasets/<dataset_name>, e.g., ~/hyperdatasets/cifar and ~/hyperdatasets/wikitext2.
Dataset files for the synthetic dataset are included directly in muir/datasets/synthetic.
Dataset files for Cifar can be downloaded directly with PyTorch.
Dataset files for WikiText-2 can be downloaded from https://www.salesforce.com/products/einstein/ai-research/the-wikitext-dependency-language-modeling-dataset/.
Dataset files for CRISPR binding prediction can be downloaded from: http://nn.cs.utexas.edu/pages/research/crispr-binding-prediction.tar.gz.
cd muir/pytorch/muir
python optimize.py --experiment_name <exp_name> --config <config_file> --device <device_id>
experiment_name is the name of the experiment and can be anything. Experiment launch time information will be appended to this name.
config is a path to the config file. For example configs, see muir/pytorch/configs.
device is the name of the device for running torch, e.g., cpu, cuda:0, cuda:1, ...
Results for the experiment will be saved to a directory with the experiments name in muir/results.
To use a new architecture, a model class can be implemented that replaces layers with hyperlayers (see muir/pytorch/models/ for examples).
Currently, layers supported for reparameterization by hypermodules are fully-connected, conv2d, conv1d, and LSTM (see muir/pytorch/layers/). These can be extended to more layer types by following the examples there.
To use a new dataset, it can be implemented to follow the interface of the examples in muir/pytorch/datasets/.
Subdirectories of this project contain the Licensing provisions made by their respective authors.
In the case of the muir/datasets/synthetic directory the author has no specific licensing requirement, however it has been shared with the permission of the author.