/ntk_matrix_completion

Primary LanguagePython

Installation

Run pip install -r requirements.txt.

Major updates

2 Feb 23

  • Same problem as 2 Jan, that directory is officially cursed. Now onwards it should be /home/mrx/projects/affinity, with data/ and output/ no longer inside ntk_matrix_completion but inside /pool001/mrx/projects/affinity on Engaging.

2 Jan 22

  • Migration over to parent directory /home/mrx/projects/binding_energies from /home/mrx/projects/matrix_completion due to sshfs problems between Engaging and deepware (might be because of Github repos and permission sharing). Hence, a lot of the file paths might change in a future commit - hopefully they will be rewired to be independent of the parent directories!
  • This directory is kept here on deepware for record keeping, but nothing will happen here after 2 Jan 22. The repos are still tied to the remote repos, but maybe they should not be.

Usage

Run bash ntk_matrix_completion/run_scripts/experiments.sh after specifying the required parameters in experiments.sh. This codebase is still undergoing changes so this section will be updated accordingly. Currently, experiments.sh contains code to run XGBoost tuning and fitting.

Code structure

config/

config contains weights for descriptors, as well as hyperparameter search spaces.

data/

data contains prior files in the prior folder, as well as ground truth files. This can also be a folder to contain output files.

features/

'features' contains prior.py for parsing prior information. Other scripts help to generate prior files from other source files and are mostly obsolete now.

models/

models contains folders for each class of models. Currently, the models are ntk, xgboost and two_tower_nn.

run_scripts/

run_scripts contains bash scripts to run multiple experiments easily, as well as python scripts that run cross-validation, feature selection, inference and so on.

  • experiments.sh is coded to run XGBoost hyperparameter tuning and model fitting and inference.
  • The rest are a little outdated and need to be refactored.

runs/

runs/ contains some output scripts from two tower NN. This should be moved to data.

tests/

tests contains scripts for testing the codebase. Currently only contains a script for testing isomer clustering.

utils/

utils contains scripts used across the codebase.

  • analysis_utilities helps in data analysis
  • hyperopt uses Hyperopt for parameter tuning
  • loss contains custom loss functions
  • non-binding contains treatment for non-binding entries
  • package_matrix helps to format the ground truth
  • path_constants contains file paths. This is in the process of being transformed into a script containing default file paths, as we move to using argparse for user-specified file paths.
  • pca contains code to do Principal Component Analysis
  • random_seed contains a record of random seeds used. This is in the process of being transformed into a script containing default seeds, as we move to using argparse for user-specified file paths.
  • utilities contains miscellaneous common code. The most important is the chunking and isomer clustering used to create data splits
  • zeolite_scrapper is used to scrap information from the IZA website. Note that the website format changed in July 2022

Old README

  • package_matrix.py: You'll want to run this with the binding.csv (https://github.com/learningmatter-mit/Zeolite-Phase-Competition/blob/main/data/binding.csv). Check out path_constants.py to see where that goes.
  • precompute_osda_priors.py: This is what you use to populate OSDA priors before running the NTK. This can take a bit if you give it tons of OSDAs. Probably best to move to cluster.
  • zeolite_scraper.py: scrapes https://america.iza-structure.org for features on all the zeolites hosted there. Built on beautiful soup and should be pretty easy to extend and scrape for additional info.
  • zeo_1.py: Holds all of the features for ZEO-1 (thanks Mingrou!)

Meat of the algorithm:

  • ntk.py: The actual implementation of the NTK & 10-fold cross validation using it.
  • prior.py: The actual meat of where priors for zeolite & ntk are formatted (before you run this you must precompute the priors or scrape them with zeolite_scraper.py)
  • utilities.py

Analysis:

  • analysis_utilities.py: Computes all the good stuff like RMSE, spearman, top_k. Written to take into account non-binding cells.
  • pca.py: This is where all the PCA plots with GETAWAY descriptors are generated. A bit messy right now, but definitely a good place to pull code from if you'd like to make some figures.

Actually user facing:

  • train.py (used to be run_method_development.py): This file contains all of the tools for prior development. I.e. it just holds a ton of functions that run 10-fold cross validation with the ground truth energy matrix.
  • run_predict.py: The code in this file was written after the algorithm/priors were looking good. It takes the complete ground_truth matrix as training and then outputs predictions for either new OSDA or new zeolite rows (depending whether or not you transpose).

Miscellaneous:

  • eigen_pro/: this folder is work in progress for getting skinny_matrix to scale well.
  • ooc_matrix_multiplication.py: also work in progress for doing the ntk on disc rather in RAM for when the matrices (specifically the kernels) get too big to handle.
  • path_constants.py: this holds all of the path constants for .pkl/.csv data files, where the program expects them to be etc.

TL;DR The two commands you'll be running for the most part: python neural_tangent_kernel/run_predict.py python neural_tangent_kernel/run_method_development.py All input & output data files are specified in neural_tangent_kernel/path_constants.py.

The files you'll care the most about at run-time:

  • neural_tangent_kernel/ntk.py: Loads the data & does all the computation
  • neural_tangent_kernel/prior.py: Prepares the priors (aka just loads priors pkl files)

The files you'll care the most about before run-time: