Unbalanced Sobolev Descent

Pytorch source code for paper

Youssef Mroueh and Mattia Rigotti, "Unbalanced Sobolev Descent", in Advances in Neural Information Processing Systems 33 (NeurIPS), Dec. 2020 [arXiv:2009.14148]

Requirements

• Python 3.6 or above
• PyTorch 1.6.0
• Numpy 1.19.1
• SciPy 1.5.2
• Matplotlib 3.3.1
• PIL 7.2.0

These can be installed using pip by running:

>> pip install -r requirements.txt

Usage

Synthetic data experiments

To reproduce the the flow simulations between synthetic distributions (Figs 1, 2, 5 and 6) first run the bash script experiments.bash, which calls the main python scripts with the appropriate paramters to reproduce the figure:

>> bash experiments.bash

The results will be saved in the folder final_outputs and will be used by the notebook plot_synthetic_data.ipynb to generate the Figures 1, 2, 5 and 6 in the paper.

Interpolation analysis of scRNA-seq data

The notebook wot_comparison.ipynb reproduces the interpolation analysis of single-cell RNA sequencing data and generates the relative plots (Figs 4 and 8 in the paper). Please, refer to the instruction in the notebook to download and prepare the data that is used.

Documents

• NeurIPS 2020 poster slides

Citation

Youssef Mroueh, Mattia Rigotti, "Unbalanced Sobolev Descent", in Advances in Neural Information Processing Systems 33 (NeurIPS), Dec. 2020 [arXiv]