ValiDyna - An evaluation framework for dynamical systems

About

This repository contains the code for my MSc thesis research at UofT and EPFL, and for the 2022 ICML SCIS Workshop paper "Understanding Generalization and Robustess of Learned Representations of Chaotic Dynamical Systems".

ValiDyna is the name given to this evaluation framework that can assess the quality of learned representations of dynamical systems. It allows to write many kinds of experiments in a multi-task setting, and includes the multi-task extension of GRU, LSTM, Transformer and N-BEATS.

Installation

0. (Pre-requisite) Install anaconda or miniconda.
1. Clone this git repository and step into it
2. Run the command conda env create -f environment.yml
   1. this will create an environment named chaos
   2. you can change the environment name in the environment.yml file before running the command
3. (For Linux), run the command export PYTHONPATH="${PYTHONPATH}:${pwd}/validyna" to add validyna to the set of libraries that can be imported

Project structure

• validyna: the source code for the project
  • main.py: the entry-point to the framework, the only part of the code to be run
  • models
    • multitask_models.py: the adaptation of N-BEATS, LSTM, GRU and Transformer to a multi-task setting
    • task_modules.py: Pytorch-Lightning modules that encapsulate specific tasks with its own training objective and metrics
    • nbeats.py: our custom implementation of the N-BEATS architecture
  • metrics.py: custom metrics
  • registry.py: contains all usable model architectures, tasks and metrics, and can be easily extended
  • data.py: a helper file with data processing and loading / saving utilities
  • plot.py: a helper file with plotting functions
• experiment_configs
  • feature_freeze: the configs for the experiment "Multi-Task Transfer-Learning" in the paper
  • few_shot: the configs for the experiment "One-shot Learning" in the paper
  • prober: the configs for a new experiment not appearing in the paper
  • default.py: the default configuration used across experiments
• scripts
  • compute_attractor_stats.py: a script that counts the number of attractors per space dimension
  • generate_data.py: a configurable script to generate synthetic chaotic data
• notebooks
  • Plot Results.ipynb: notebook downloading metrics from wandb and generating plots
• tests: directory with the same structure as validyna containing tests per file

Running experiments and writing new ones

Current experiment configurations are in the folder experiment_configs.

To run an experiment, execute the command: python validyna/main.py --cfg=path/to/experiment_config.py (make sure the third installation step is complete).

To add a new experiment config, check the documentation of run_experiment and run_model_training in main.py, as well as ChunkMultiTaskDataset in data.py.

Paper and citation

You can find the ValiDyna paper here: https://openreview.net/forum?id=sJg-BRQcLxr

If you use ValiDyna in your work, please cite us using the following bibtex:

@inproceedings{streit2022understanding,
    title={Understanding Generalization and Robustess of Learned Representations of Chaotic Dynamical Systems},
    author={Lu{\~a} Streit and Vikram Voleti and Tegan Maharaj},
    booktitle={ICML 2022: Workshop on Spurious Correlations, Invariance and Stability},
    year={2022},
    url={https://openreview.net/forum?id=sJg-BRQcLxr}
}