Thank you for your work on the efficient way to predict the ML method for the molecular system.

But, I couldn't reproduce the paper.

I found that Jarvis summarizes QM9 datasets with normalization, and I issued in jarvis.

I tested ALIGNN, but cannot reproduce it for the unnormalized QM9 datasets. Only the normalized QM9 dataset provided by Jarvis works to reproduce prediction values in paper.

Dear All,

I successfully trained a model with ALIGNN. There is "pretrained.py" script that import some models according to datasets.

My question is how I can use my trained model to predict a structure? Do I have to modify the "pretrained.py" script to import my model or can I do something like this:

python alignn/pretrained.py --model_name <my_output_pt_file> --file_format poscar --file_path /path/to/sample_file

Thanks for the help.

Dear All,

First of all, thanks for creating ALIGNN tool.

I am trying to train a model with OMDB dataset to obtain a bandgap prediction. The dataset containes xyz files of molecules and bandgap values of them. It is also included in JARVIS documentation: https://jarvis-tools.readthedocs.io/en/master/databases.html

I am following the README file on ALIGNN page. I generated my xyz samples from the dataset as it follows without any problem:

` from jarvis.db.figshare import data as jdata from jarvis.core.atoms import Atoms

omdbset = jdata("omdb") prop = "bandgap"

max_samples = 12500 f = open("id_prop.csv", "w") count = 0 for i in omdbset: atoms = Atoms.from_dict(i["atoms"]) cod_id = i["cod_id"] xyz_name = "OMDB-" + cod_id + ".xyz" target = i[prop] if target != "na": atoms.write_xyz(xyz_name) f.write("%s,%6f\n" % (xyz_name, target)) count += 1 if count == max_samples: break f.close() `

I used the config.json file like you did it in QM9 training. Just the following 2 lines are different:

"dataset": "omdb", "target": "bandgap",

When i tried to run the code like this, it gave the following errors:

python /home/fsysadmin/alignn/alignn/train_folder.py --root_dir "/home/fsysadmin/alignn/omdb_tests/prep_data" --config "/home/fsysadmin/alignn/omdb_tests/prep_data/config.json" --file_format xyz --output_dir=/home/fsysadmin/alignn/omdb_tests/results

Using backend: pytorch Check 1 validation error for TrainingConfig dataset unexpected value; permitted: 'dft_3d', 'jdft_3d-8-18-2021', 'dft_2d', 'megnet', 'megnet2', 'mp_3d_2020', 'qm9', 'qm9_dgl', 'qm9_std_jctc', 'user_data', 'oqmd_3d_no_cfid', 'edos_up', 'edos_pdos', 'qmof', 'hmof', 'hpov', 'pdbbind', 'pdbbind_core' (type=value_error.const; given=omdb; permitted=('dft_3d', 'jdft_3d-8-18-2021', 'dft_2d', 'megnet', 'megnet2', 'mp_3d_2020', 'qm9', 'qm9_dgl', 'qm9_std_jctc', 'user_data', 'oqmd_3d_no_cfid', 'edos_up', 'edos_pdos', 'qmof', 'hmof', 'hpov', 'pdbbind', 'pdbbind_core')) Traceback (most recent call last): File "/home/fsysadmin/alignn/alignn/train_folder.py", line 194, in <module> train_for_folder( File "/home/fsysadmin/alignn/alignn/train_folder.py", line 80, in train_for_folder config.keep_data_order = keep_data_order AttributeError: 'dict' object has no attribute 'keep_data_order'

As i understand from the error output, OMDB dataset is not included into ALIGNN package. Giving the full path of OMDB tar file did not also solve the problem.

How can i include OMDB dataset? There are some scripts that import dataset in alignn repo such as train_all_qm9_jctc.py. Maybe this scripts can be modified to include OMDB.

I appreciate your help.

Best regards,

Is the JARVIS-ALIGNN web interface accessible via Python API? I'd like to get predictions for a few dozen POSCAR files without pasting them all in manually.

let's do this in cross validation to address the stability question

strategy: 5x shuffle-split validation scheme to keep things simple. schedule runs using ray tune grid_search

report results for jarvis-55k formation energy and band gap targets.

• [x] jarvis-55k formation energy
• [x] jarvis-55k band gap
• [x] publication-quality plots
• [x] integrate into manuscript

• Fix pytorch version to match torchvision that can be used in atomvision package.
• https://github.com/pytorch/vision/blob/cb60e97a778ad6626980b67c456984e5fcf1d507/README.rst#L22

Also required adding max_neighbors as a parser argument Also involved adding an optional field for extra config parameters in the "all models" list

checking through atomwise training loop and force computation

I think we should probably consolidate around the idea of storing target tensors and output tensors in dictionaries. then we can generate a custom loss function and go back to using ignite, and hopefully remove some of the code complexity

Improve sections of train_dgl that write the last epoch's train and validation predictions to files. Also improve the write handling such that batch sizes of 1 are concatenated to files without requiring special case.

Works only with PanayotisManganaris:pm/alignn-patch #261 on jarvis.

Hello everyone :)

Thanks for great tool! I have found a small error in the "alignn/train.py" script while I was trying to use it for regression.

At line 866 we have: `

if config.n_early_stopping is not None:
    if classification:
        my_metrics = "accuracy"
    else:
        my_metrics = "mae"

    def default_score_fn(engine):
        score = engine.state.metrics[my_metrics]
        return score

    es_handler = EarlyStopping(
        patience=config.n_early_stopping,
        score_function=default_score_fn,
        trainer=trainer,
    )

` The problem is that, as stated in the documentation of ignite, "An improvement is considered if the score is higher." In other words, EarlyStopping only checks for increases of a benefit/reward function to be maximized, like "accuracy". A decreasing cost function would trigger earlyStopping after "patience" number of epochs, even if the cost is still decreasing.

To monitor the decrease of a cost function, e.g. "mae", the default_score_fn() could return the negative value of the cost metric, as shown in the ignite documentation example here: https://pytorch.org/ignite/generated/ignite.handlers.early_stopping.EarlyStopping.html

All the best

Dear All,

I train ALIGNN with cif files. To improve the performance, I tried to augment my cif files with AugLiChem library. Here is the snippet from the original file and the augmented cif file:

Original file:

data_image0
_chemical_formula_structural       H10C14S2N2O2
_chemical_formula_sum              "H10 C14 S2 N2 O2"
_cell_length_a       4.3258
_cell_length_b       8.982
_cell_length_c       8.4721
_cell_angle_alpha    90
_cell_angle_beta     90.594
_cell_angle_gamma    90

_space_group_name_H-M_alt    "P 1"
_space_group_IT_number       1

loop_
  _space_group_symop_operation_xyz
  'x, y, z'

loop_
  _atom_site_type_symbol
  _atom_site_label
  _atom_site_symmetry_multiplicity
  _atom_site_fract_x
  _atom_site_fract_y
  _atom_site_fract_z
  _atom_site_occupancy
  H   H1        1.0  0.83700  0.83300  0.55800  1.0000
  H   H2        1.0  0.16300  0.33300  0.44200  1.0000
.
.
.
(continues)

Augmented file:

# generated using pymatgen
data_H5C7SNO
_symmetry_space_group_name_H-M   'P 1'
_cell_length_a   4.32580000
_cell_length_b   8.98200000
_cell_length_c   8.47210000
_cell_angle_alpha   90.00000000
_cell_angle_beta   90.59400000
_cell_angle_gamma   90.00000000
_symmetry_Int_Tables_number   1
_chemical_formula_structural   H5C7SNO
_chemical_formula_sum   'H10 C14 S2 N2 O2'
_cell_volume   329.16012678
_cell_formula_units_Z   2
loop_
 _symmetry_equiv_pos_site_id
 _symmetry_equiv_pos_as_xyz
  1  'x, y, z'
loop_
 _atom_site_type_symbol
 _atom_site_label
 _atom_site_symmetry_multiplicity
 _atom_site_fract_x
 _atom_site_fract_y
 _atom_site_fract_z
 _atom_site_occupancy
  H  H0  1  0.83420779  0.82961480  0.54722879  1.0
  H  H1  1  0.15729856  0.32521300  0.43626670  1.0
.
.
.
(continues)

ALIGNN works fine with original cif files but whenever I try to train it with augmented file, I encounter the following error:

Using backend: pytorch
Traceback (most recent call last):
  File "/raid/apps/alignn/2021/bin/train_folder.py", line 195, in <module>
    train_for_folder(
  File "/raid/apps/alignn/2021/bin/train_folder.py", line 103, in train_for_folder
    atoms = Atoms.from_cif(file_path)
  File "/raid/apps/alignn/2021/lib/python3.8/site-packages/jarvis/core/atoms.py", line 537, in from_cif
    cif_atoms = cif_atoms.get_primitive_atoms
  File "/raid/apps/alignn/2021/lib/python3.8/site-packages/jarvis/core/atoms.py", line 710, in get_primitive_atoms
    return Spacegroup3D(self).primitive_atoms
  File "/raid/apps/alignn/2021/lib/python3.8/site-packages/jarvis/analysis/structure/spacegroup.py", line 240, in primitive_atoms
    lattice, scaled_positions, numbers = spglib.find_primitive(
TypeError: cannot unpack non-iterable NoneType object

I can not see a problem in augmented files. Do you have any suggestions?

Best regards,

Dear All,

I am using ALIGNN model to train OMDB dataset and trying to improve results by adjusting hyperparameters. But I have not achieved good results yet (one of the reasons is that my test molecules are a little bigger than the dataset)

In README.md file, OMDB results say "coming soon". Do you have any train results for this? I would like to compare your result and test the trained model on my own molecules.

Best regards,

Hi,

Thanks for this great library.

I think I noticed a slight inconsistency in the code from the paper. The paper states that the ALIGNN layer first performs the edge-gated graph convolution on the line graph to update the pair and triplet features, and then the pair features are passed as edges to the atomistic/direct graph.

However, when I look at alignn.models.alignn.ALIGNNConv.forward, I see that the edge-gated graph convolution is actually applied on the atomistic/direct graph first, and then the updated pair features are passed as nodes to the line graph. Am I understanding this correctly?

Hi @bdecost,

When I wrote the following code based on your code to print the line graph, an incomprehensible part appeared.

In the code below, among (node_i, node_j) based on ij_pair and (nodej_, nodek_) based on jk_pair, I think that (node_j) and (nodej_) should have the same node number, but there are cases where they do not match as a result of executing the code below.

When I referred to your paper, I understood that the angle of the line graph is made based on node_i, node_j, and node_k. Why is the shared node_j not the same?

############################################ CIF file (mp-2500.cif) '''generated using pymatgen''' data_AlCu _symmetry_space_group_name_H-M 'P 1' _cell_length_a 6.37716407 _cell_length_b 6.37716407 _cell_length_c 6.92031335 _cell_angle_alpha 57.14549155 _cell_angle_beta 57.14549155 _cell_angle_gamma 37.46229262 _symmetry_Int_Tables_number 1 _chemical_formula_structural AlCu _chemical_formula_sum 'Al5 Cu5' _cell_volume 140.31041575 cell_formula_units_Z 5 loop _symmetry_equiv_pos_site_id symmetry_equiv_pos_as_xyz 1 'x, y, z' loop _atom_site_type_symbol _atom_site_label _atom_site_symmetry_multiplicity _atom_site_fract_x _atom_site_fract_y _atom_site_fract_z _atom_site_occupancy Al Al0 1 0.50000000 0.50000000 0.50000000 1 Al Al1 1 0.15622000 0.15622000 0.53856900 1 Al Al2 1 0.84378000 0.84378000 0.46143100 1 Al Al3 1 0.37823100 0.37823100 0.00427500 1 Al Al4 1 0.62176900 0.62176900 0.99572500 1 Cu Cu5 1 0.00000000 0.00000000 0.00000000 1 Cu Cu6 1 0.25794700 0.25794700 0.75941600 1 Cu Cu7 1 0.74205300 0.74205300 0.24058400 1 Cu Cu8 1 0.10895200 0.10895200 0.22813800 1 Cu Cu9 1 0.89104800 0.89104800 0.77186200 1

############################################# Code import os import dgl import numpy as np

import torch

from jarvis.core.atoms import Atoms from jarvis.core.graphs import Graph

from torch_geometric.data import InMemoryDataset, Data, Batch from torch_geometric.utils.convert import from_networkx

raw_path = './mp-2500.cif' crystal = Atoms.from_cif(raw_path, use_cif2cell=False) coords = crystal.cart_coords graph = Graph.atom_dgl_multigraph(crystal, cutoff=8.0, atom_features='cgcnn',max_neighbors=12, compute_line_graph=True, use_canonize=False) for i in [0,1]: '''Atom-Bond Graph''' if i==0: g = from_networkx(dgl.DGLGraph.to_networkx(graph[i], node_attrs=['atom_features'], edge_attrs=['r'])) x = torch.tensor([x.detach().numpy() for x in g.atom_features]) z = torch.tensor(crystal.atomic_numbers) pos = torch.tensor(coords, dtype=torch.float) edge_id = g.id edge_pos = torch.tensor([x.detach().numpy() for x in g.r]) edge_index = g.edge_index edge_distance = torch.tensor(np.linalg.norm(graph[i].edata['r'], axis=1)) ab_g = Data(x=x, z=z, pos=pos, edge_id=edge_id, edge_index=edge_index, edge_distance=edge_distance, edge_pos=edge_pos, idx=n) '''Line Graph''' if i==1: g = from_networkx(dgl.DGLGraph.to_networkx(graph[i], node_attrs=['r'], edge_attrs=['h'])) x = torch.tensor(np.linalg.norm(graph[i].ndata['r'], axis=1)) pos = torch.tensor([x.detach().numpy() for x in g.r]) edge_id = g.id edge_index = g.edge_index edge_angle = g.h ba_g = Data(x=x, pos=pos, edge_id=edge_id, edge_index=edge_index, edge_angle=edge_angle, idx=n) dataset = [ab_g, ba_g]

'''dataset[1] = Line Graph''' '''dataset[0] = Atom-Bond Graph''' ij_pair = dataset[1].edge_index[0] jk_pair = dataset[1].edge_index[1] node_i = dataset[0].edge_index[0][ij_pair] node_j = dataset[0].edge_index[1][ij_pair] nodej_ = dataset[0].edge_index[0][jk_pair] nodek_ = dataset[0].edge_index[1][jk_pair]

################################################################# Result node_i[0:10], node_j[0:10], nodej_[0:10], nodek_[0:10]