v0.3.0

[cw-tan]

[0.3.0]

Added

• Hyperparameters tensors_mixing_mode, weight_individual_irreps, typexbasis_mode, tensor_track_weight_init, self_edge_tensor_product
• New Bessel basis with hyperparameters num_bessels_per_basis, bessel_frequency_cutoff, per_edge_type_cutoff

Changed

• Resnet normalization scheme
• [Breaking] latent_resnet: false enables skip connections
• [Breaking] Default for norm_basis_mean_shift changed from True -> False
• [Breaking] Rename env_embed_mul -> num_tensor_features
• [Breaking] Replaced nonscalars_include_parity with tensor_track_allowed_irreps

---

This discussion was created from the release v0.3.0.

[Enry99]

Is this the major code update stated in the README, or will there be another release soon?

[cw-tan]

There's more to come, this is just to get the last allegro version that works with the older nequip infrastructure out version tagged and on PyPI.

[hkqiu]

Is there any colab notebook available like before for learning？

[cw-tan]

Alongside the allegro v0.4.0 update, there is a corresponding nequip v0.7.0 update. There is a Colab tutorial in the nequip README.md, where one can learn how to use the NequIP framework. Allegro is a specific model that is used in the NequIP framework, and there's a minimal example to show how one can modify the model portion of the config file to use Allegro models in the NequIP framework here.

[musadu2]

Hey everybody, I have a question about num_bessels_per_basis hyper parameter!

I used run allegro 0.2.0 and now running 0.3.0. There is this new parameter that I need to defined but I want my current trainings to be consistent with my previous runs.

What should I set the num_bessels_per_basis parameter?

I would be very glad if you can answer this question.

[cw-tan]

It refers to the number of Bessel basis functions, usually 8 is a reasonable choice.

[musadu2]

Thanks for your answer!! I have a question about the Allegro training process. are the the configurations are wrapped internally in Allegro mlip? I mean that whether Allegro automatically wraps atomic configurations internally using periodic boundary conditions (PBCs) if PBCs are set as "T T T"?

To understand this question by myself, I tested a sample Allegro model on same wrapped/unwrapped configurations.
The test results on unwrapped configurations:
--- Final result: ---
f_mae = 0.229882
f_rmse = 0.337401
H_f_mae = 0.229882
psavg_f_mae = 0.229882
H_f_rmse = 0.337401
psavg_f_rmse = 0.337401
e_mae = 2.113169
e/N_mae = 0.022012
e_rmse = 2.560476
e/N_rmse = 0.026672

The test results on wrapped configurations:
--- Final result: ---
f_mae = 0.229882
f_rmse = 0.337400
H_f_mae = 0.229882
psavg_f_mae = 0.229882
H_f_rmse = 0.337400
psavg_f_rmse = 0.337400
e_mae = 2.113181
e/N_mae = 0.022012
e_rmse = 2.560488
e/N_rmse = 0.026672

from here, I see that the differences between energy and force RMSE are on the order of ~0.000001, and I think that is a negligible difference.

I would be very glad if you can answer this question.

[musadu2]

I have been also reading the nequip source code, and here is what I found:

I found the wrapping python code inside nequip
(https://erg-illinois.slack.com/archives/D05NYBYAB2R/p1750782153463409)
it automatically wraps the atoms in the code
(https://erg-illinois.slack.com/archives/D05NYBYAB2R/p1750782277970989)
https://github.com/mir-group/nequip/blob/main/nequip/data/AtomicDataDict.py
at the beginning of this code Wrapping of atoms across periodic boundaries is encoded in the edge_cell_shift tensor
https://github.com/mir-group/nequip/blob/main/nequip/data/datamodule/_base_datamodule.py
in this code:
if "collate_fn" not in dataloader_dict:
dataloader_dict["collate_fn"] = {
"target": "nequip.data.datamodule._base_datamodule._default_collate_fn_factory"
}
When NequIP builds a PyTorch DataLoader, it uses a custom collate_fn to assemble a batch.
The default collate_fn is _default_collate_fn_factory, which returns a function that callsAtomicDataDict.batched_from_list
AtomicDataDict.batched_from_list(...) respects edge_index, edge_cell_shift, cell, and other keys needed to track periodic images.

[Linux-cpp-lisp]

Hi @musadu2 , can you please open a new GitHub discussions thread for this on the NequIP repo? Thanks!

[musadu2]

I have a question about the Allegro training process. are the the configurations are wrapped internally in Allegro mlip? I mean that whether Allegro automatically wraps atomic configurations internally using periodic boundary conditions (PBCs) if PBCs are set as "T T T"?

To understand this question by myself, I tested a sample Allegro model on same wrapped/unwrapped configurations.
The test results on unwrapped configurations:
--- Final result: ---
f_mae = 0.229882
f_rmse = 0.337401
H_f_mae = 0.229882
psavg_f_mae = 0.229882
H_f_rmse = 0.337401
psavg_f_rmse = 0.337401
e_mae = 2.113169
e/N_mae = 0.022012
e_rmse = 2.560476
e/N_rmse = 0.026672

The test results on wrapped configurations:
--- Final result: ---
f_mae = 0.229882
f_rmse = 0.337400
H_f_mae = 0.229882
psavg_f_mae = 0.229882
H_f_rmse = 0.337400
psavg_f_rmse = 0.337400
e_mae = 2.113181
e/N_mae = 0.022012
e_rmse = 2.560488
e/N_rmse = 0.026672

from here, I see that the differences between energy and force RMSE are on the order of ~0.000001, and I think that is a negligible difference.
I have been also reading the nequip source code, and here is what I found:

I found the wrapping python code inside nequip
(https://erg-illinois.slack.com/archives/D05NYBYAB2R/p1750782153463409)
it automatically wraps the atoms in the code
(https://erg-illinois.slack.com/archives/D05NYBYAB2R/p1750782277970989)
https://github.com/mir-group/nequip/blob/main/nequip/data/AtomicDataDict.py
at the beginning of this code Wrapping of atoms across periodic boundaries is encoded in the edge_cell_shift tensor
https://github.com/mir-group/nequip/blob/main/nequip/data/datamodule/_base_datamodule.py
in this code:
if "collate_fn" not in dataloader_dict:
dataloader_dict["collate_fn"] = {
"target": "nequip.data.datamodule._base_datamodule._default_collate_fn_factory"
}
When NequIP builds a PyTorch DataLoader, it uses a custom collate_fn to assemble a batch.
The default collate_fn is _default_collate_fn_factory, which returns a function that callsAtomicDataDict.batched_from_list
AtomicDataDict.batched_from_list(...) respects edge_index, edge_cell_shift, cell, and other keys needed to track periodic images.

I would be very glad if you can answer this question.

[cw-tan]

Our models account for PBCs if provided -- see https://github.com/mir-group/nequip/blob/9eaae58cd16016d43b2999e231c9d27c21f7a0e0/nequip/nn/utils.py#L68

Both Allegro and NequIP use that utility function for computing edge vectors, which accounts for periodic images through edge_cell_shifts, which will be provided if there are PBCs.

Echoing @Linux-cpp-lisp's comment -- it's good practice to start new discussions for conceptually different topics. Also, questions should be directed to discussions, not issues. Issues are meant for bug reports, feature requests, and the like.