Changelog¶
All notable changes to metatrain
are documented here, following the keep a
changelog format. This project follows
Semantic Versioning.
Unreleased¶
Added¶
Additional logs to the checkpoints, model and the output dirs at the end of training
When downloading checkpoints and models from Hugging Face, the files will be cached locally and re-used.
extra_data
is now a valid section in theoptions.yaml
file, allowing users to add custom data to the training set. The data is contained in the dataloader and can be used in custom loss functions or models.mtt eval
can be used to evaluate models on aDiskDataset
.
Version 2025.8 - 2025-06-11¶
Changed¶
Checkpoints for fine-tuning files are now passed from the
options.yaml
.
Version 2025.7 - 2025-05-27¶
Changed¶
Metatrain is now built on top of
metatomic
instead ofmetatensor.torch.atomistic
. Please refer to https://docs.metatensor.org/metatomic/ to find how to use the new models.
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