Changelog¶
All notable changes to metatrain
are documented here, following the keep a
changelog format. This project follows
Semantic Versioning.
Unreleased¶
Version 2025.6 - 2025-04-28¶
Fixed¶
PET
can now evaluate on single-atom structures without crashingThe metatrain dataloader doesn’t load all batches ahead of each epoch anymore
Added¶
NanoPET
andPET
can now train on non-conservative stressesUsers can now choose the name of the extension directory in
mtt train
andmtt export
via the--extensions
(or-e
) optionUpdate to
metatensor-torch-0.7.6
, adding support for torch 2.7PET
now supports gradient clipping as a new training hyperparameter
Changed¶
Training and exporting models without extensions will no longer lead to the creation of an empty directory for the extensions
The SOAP-BPNN model now uses
torch-spex
instead offeatomic
as its SOAP backendPET
from the previous version is now deprecated and accessible asdeprecated.pet
, while the oldNativePET
(experimental.nativepet
) is now calledPET
(pet
from training option files)The Angstrom character is now represented as
A
and notÅ
in the training logs
Version 2025.5 - 2025-04-13¶
Fixed¶
Fix more composition model issues
Added¶
Update to
metatensor-torch-0.7.5
to allow training onnon_conservative_forces
andnon_conservative_stress
targetsAdd
NativePET
as a readable, efficient, backward-compatible PET implementationAdded Wandb logger
Save loss history in a
.csv
file
Version 2025.4 - 2025-03-29¶
Changed¶
upgraded to
metatensor.torch
0.7.4, which gives access to batched ASE evaluation
Version 2025.3 - 2025-03-25¶
Fixed¶
Fixed a bug in the composition model, affecting SOAP-BPNN and nanoPET
Changed¶
metatrain.util.io.load_model()
does not copy a remote model to the current directory.
Version 2025.2 - 2025-03-11¶
Added¶
Implement a long-range featurizer as a utility for all models
Speed up system preparation
Changed¶
Remove biases in SOAP-BPNN’s linear layers
Fixed¶
Fix NanoPET multi-GPU error message
Fix
device
for fixed composition weights
Version 2025.1 - 2025-02-20¶
Added¶
Support for Python 3.13 and
ase
>= 3.23
Fixed¶
Some irrelevant autograd warnings
Version 2025.0 - 2025-02-19¶
Added¶
First release outside of the lab