v1.3.0 - Into the multi-GPU-niverse

New features

• Multi-GPU inference: Models can now make predictions on an arbitrary number of GPUs

• Multi-GPU training: Models can now be trained on an arbitrary number of GPUs

• MALA now works with 2D materials, i.e., any system which is only periodic in two dimensions

• Bispectrum descriptor calculation now possible in python

  • This is route is significantly slower than LAMMPS, but can be helpful for developers who want to test the entire MALA modeling workflow without installing LAMMPS

• Logging for network training has been overhauled and now allows for the logging of multiple metrics

• (EXPERIMENTAL) Implementation of a mutual information based metric to replace/complement the ACSD

• (EXPERIMENTAL) Implementation of a class for LDOS alignment to a reference energy value; this can be useful for models across multiple mass densities

Changes to API/user experience

• New parallelization parameters available:
  • use_lammps - enable/disable LAMMPS (enabled by default, recommended for optimal performance, will automatically be disabled if no LAMMPS is found on the machine)
  • use_atomic_density_formula - enable the use of total energy evaluation based on a Gaussian representation (enabled if LAMMPS and GPU are enabled, recommended for optimal performance, details can be found in our paper on size transfer)
  • use_ddp - enable/disable DDP, i.e., Pytorch's distributed training scheme (disabled by default)
• Multiple LAMMPS/QE calculations can now be run in one directory
  • Prior to this version, doing so would lead to problems due to the file based nature of these interfaces
  • This allows for multiple simultaneous inferences in the same folder
• Class SNAP and all associated options are deprecated, use Bispectrum and associated options instead
• Default units for reading from .cube files are now set to units commonly used within Quantum ESPRESSO, this should make it easier to avoid inconsistencies in data sampling
• ASE calculator MALA now reads models with load_run() instead of load_model which is more consistent with the rest of MALA
• Error reporting with the Tester class has been improved, all errors and energy values reported there are now consistently given in meV/atom
• MALA calculators (LDOS, density, DOS) now also read energy contributions and forces from Quantum ESPRESSO output files, these can be accessed via properties

Fixes

• Updated various performance/accessibility issues of CI/CD
• Fixed compatability with newer Optuna versions
• Added missing docstrings
• Fixed shuffling interface, arbitrary numbers of shuffled snapshots can now be created without loss of information
• Fixed inconsistency of density dimensions when using directly from cube file
• Fixed error when using GPU graphs with arbitrary batch sizes

v1.2.1 - Minor bufixes

This release fixes some minor issues and bugs, and updates some of the meta information. It also serves as a point of reference for an upcoming scientific work.

Change notes:

• Updated MALA logos
• Updated Tester class to also give Kohn-Sham energies alongside LDOS calculations
• Updated CITATION.cff file to reflect new team members and scientific supervisors
• Fixed bug that would crash models trained with horovod when loaded for inference without horovod
• Fixed bug that would crash training when using Optuna+MPI for hyperparameter optimization (GPU compute graph usage had not been properly adaptable for this scenario)
• Deactivated pytorch profiling by default, can still be manually enabled

v1.2.0 - GPU and you

New features

• Production-ready inference options
  • Full inference (from ionic configuration to observables) on either a single GPU or distributed across multiple CPU (multi-GPU support still in development)
  • Access to (volumetric) observables within seconds
• Fast training speeds due to optimal GPU usage
• Training on large data sets through improved lazy-loading functionalitites and data shuffling routines
• Fast hyperparameter optimization through distributed optimizers (optuna) and training-free surrogate metrics (NASWOT/ACSD)
• Easy-to-use interface through single Parameters object for reproducibolity and modular design
• Internal caching system for intermediate quantities (e.g. DOS, density, band energy) for improved performance
• Experimental features for advanced users:
  • MinterPy: Polynomial interpolation based descriptors
  • OpenPMD
  • OF-DFT-MD interface to create initial configurations for ML based sampling

Change notes:

• Full (serial) GPU inference added
• MALA now operates on FP32
• Added functionality for data shuffling
• Added functionality for cached lazy loading
• Improved GPU usage during training
• Added convencience functions, e.g., for ACSD analysis
• Fixed several bugs across the code
• Overhauled documentation

v1.1.0 - (very late) Spring cleaning

Features

• Parallel preprocessing, network training and model inference
• Distributed hyperparameter optimization (Optuna) and distributed training-free network architecture optimization (NASWOT)
• Reproducibility through single Parameters object, easy interface to JSON for automated sweeps
• Internal caching system for intermediate quantities (e.g. DOS, density, band energy) for improved performance
• Modular design
• OF-DFT-MD interface to create initial configurations for ML based sampling

Change notes:

• MALA now operates internally in Angstrom consistently
  • Volumetric data that has been created with MALA v1.0.0 can still be used, but unit conversion has to be added to the scripts in question
• Implemented caching functionality
  • The old post-processing API is still fully functional, but will not use the caching functions; instead, MALA now has a more streamlined API tying calculators to data
• More flexible data conversion methods
• Improved Optuna distribution scheme
• Implemented parallel total energy inference
• Reduced import time for MALA module
• Several smaller bugfixes

v1.0.0 - First major release (PyPI version)

Features

• Preprocessing of QE data using LAMMPS interface and LDOS parser (parallel via MPI)
• Networks can be created and trained using pytorch (arallel via horovod)
• Hyperparameter optimization using optuna, orthogonal array tuning and neural architecture search without training (NASWOT) supported
  • optuna interface supports distributed runs and NASWOT can be run in parallel via MPI
• Postprocessing using QE total energy module (available as separate repository)
  • Network inference parallel up to the total energy calculation, which currently is still serial.
• Reproducibility through single Parameters object, easy interface to JSON for automated sweeps
• Modular design

Change notes:

• full integration of Sandia ML-DFT code into MALA (network architectures, misc code still open)
• Parallelization of routines:
  • Preprocessing (both SNAP calculation and LDOS parsing)
  • Network training (via horovod)
  • Network inference (except for total energy)
• Technical improvements:
  • Default parameter interface is now JSON based
  • internal refactoring

v1.0.0 - First major release

Features

• Preprocessing of QE data using LAMMPS interface and LDOS parser (parallel via MPI)
• Networks can be created and trained using pytorch (arallel via horovod)
• Hyperparameter optimization using optuna, orthogonal array tuning and neural architecture search without training (NASWOT) supported
  • optuna interface supports distributed runs and NASWOT can be run in parallel via MPI
• Postprocessing using QE total energy module (available as separate repository)
  • Network inference parallel up to the total energy calculation, which currently is still serial.
• Reproducibility through single Parameters object, easy interface to JSON for automated sweeps
• Modular design

Change notes:

• full integration of Sandia ML-DFT code into MALA (network architectures, misc code still open)
• Parallelization of routines:
  • Preprocessing (both SNAP calculation and LDOS parsing)
  • Network training (via horovod)
  • Network inference (except for total energy)
• Technical improvements:
  • Default parameter interface is now JSON based
  • internal refactoring

v0.2.0 - Regular Update

Regular update of MALA. This release mostly updates the hyperparameter optimization capabilites of MALA and fixes some minor bugs. Changelog:

• Fixed installation instructions and OAT part of installation
• Improved and added examples; made LDOS based examples runnable
• Fixed direct string concat for file interaction and replaced it with path functions
• Improved optuna hyperparameter optimization (ensemble objectives, band energy as validation loss, distributed optimization, performance study)
• Improved OAT and NASWOT implementation
• Fixed several things regarding documentation and citation
• Added check ensuring that QE-MD generated input files adhere to PBC
• Implemented visualization via tensorbard
• Stylistic improvements (import ordering fixed, TODOs converted to issues or resolved, replaced unncesseary get_data_repo() function
• Added bump version
• Set up mirror to casus org and fix pipeline deployment issues when working from forks

Test data repository version: v1.1.0

v0.1.0 - Accelerating Finite-Temperature DFT with DNN

First alpha release of MALA. This code accompanies the publication of the same name (https://doi.org/10.1103/PhysRevB.104.035120).

Features:

• Preprocessing of QE data using LAMMPS interface and parsers
• Networks can be created and trained using pytorch
• Hyperparameter optimization using optuna
  • experimental: orthogonal array tuning and neural architecture search without training supported
• Postprocessing using QE total energy module (available as separate repository)

Test data repository version: v0.1.0

v0.0.2

Added code from Sandia National Laboratories and Oak Ridge National Laboratory. Code developments will be merged beginning now.

v0.0.1

Current features:

• Preprocessing of QE data using LAMMPS interface and parsers
• Networks can be created and trained using pytorch
• Hyperparameter optimization using optuna
  • experimental: orthogonal array tuning and neural architecture search without training supported
• Postprocessing using QE total energy module