[Feature Request]: Which version do the CHGNet tutorials correspond to?

[J-You]

Email (Optional)

[email protected]

Problem

Could you please clarify which version the CHGNet tutorials correspond to? I encountered errors while running the tutorials with the latest version (0.3.8). Pymatgen version (2024.8.9)

import torch
torch.cuda.is_available()
>>> True

from chgnet.model import CHGNet
import numpy as np
from pymatgen.core import Structure

np.set_printoptions(precision=4, suppress=True)

chgnet = CHGNet.load(model_name='0.3.0', use_device='cpu')
# Alternatively you can read your own model
# chgnet = CHGNet.from_file(model_path)

structure = Structure.from_file("LiMnO2.cif")
structure
>>> Structure Summary
Lattice
    abc : 2.868779 4.634475 5.832507
 angles : 90.0 90.0 90.0
 volume : 77.5448402400077
      A : 2.868779 0.0 1.7566205099055724e-16
      B : 7.452804216860131e-16 4.634475 2.837797487239215e-16
      C : 0.0 0.0 5.832507
    pbc : True True True
PeriodicSite: Li0 (Li+) (1.434, 2.317, 2.215) [0.5, 0.5, 0.3798]
PeriodicSite: Li1 (Li+) (0.0, 0.0, 3.618) [0.0, 0.0, 0.6202]
PeriodicSite: Mn2 (Mn3+) (1.434, 2.317, 5.035) [0.5, 0.5, 0.8633]
PeriodicSite: Mn3 (Mn3+) (0.0, 0.0, 0.7976) [0.0, 0.0, 0.1367]
PeriodicSite: O4 (O2-) (1.434, 0.0, 2.105) [0.5, 0.0, 0.3608]
PeriodicSite: O5 (O2-) (3.726e-16, 2.317, 0.5746) [0.0, 0.5, 0.09851]
PeriodicSite: O6 (O2-) (1.434, 0.0, 5.258) [0.5, 0.0, 0.9015]
PeriodicSite: O7 (O2-) (3.726e-16, 2.317, 3.728) [0.0, 0.5, 0.6392]

prediction = chgnet.predict_structure(structure=structure,task='ef')

ValueError                                Traceback (most recent call last)

Cell In[7], line 1
----> 1 prediction = chgnet.predict_structure(structure=structure,task='ef')


File ~\anaconda3\lib\site-packages\chgnet\model\model.py:572, in CHGNet.predict_structure(self, structure, task, return_site_energies, return_atom_feas, return_crystal_feas, batch_size)
    568     raise ValueError("graph_converter cannot be None!")
    570 structures = [structure] if isinstance(structure, Structure) else structure
--> 572 graphs = [self.graph_converter(struct) for struct in structures]
    573 return self.predict_graph(
    574     graphs,
    575     task=task,
   (...)
    579     batch_size=batch_size,
    580 )


File ~\anaconda3\lib\site-packages\chgnet\model\model.py:572, in <listcomp>(.0)
    568     raise ValueError("graph_converter cannot be None!")
    570 structures = [structure] if isinstance(structure, Structure) else structure
--> 572 graphs = [self.graph_converter(struct) for struct in structures]
    573 return self.predict_graph(
    574     graphs,
    575     task=task,
   (...)
    579     batch_size=batch_size,
    580 )


File ~\anaconda3\lib\site-packages\torch\nn\modules\module.py:1553, in Module._wrapped_call_impl(self, *args, **kwargs)
   1551     return self._compiled_call_impl(*args, **kwargs)  # type: ignore[misc]
   1552 else:
-> 1553     return self._call_impl(*args, **kwargs)


File ~\anaconda3\lib\site-packages\torch\nn\modules\module.py:1562, in Module._call_impl(self, *args, **kwargs)
   1557 # If we don't have any hooks, we want to skip the rest of the logic in
   1558 # this function, and just call forward.
   1559 if not (self._backward_hooks or self._backward_pre_hooks or self._forward_hooks or self._forward_pre_hooks
   1560         or _global_backward_pre_hooks or _global_backward_hooks
   1561         or _global_forward_hooks or _global_forward_pre_hooks):
-> 1562     return forward_call(*args, **kwargs)
   1564 try:
   1565     result = None


File ~\anaconda3\lib\site-packages\chgnet\graph\converter.py:129, in CrystalGraphConverter.forward(self, structure, graph_id, mp_id)
    123 atom_frac_coord = torch.tensor(
    124     structure.frac_coords, dtype=datatype, requires_grad=True
    125 )
    126 lattice = torch.tensor(
    127     structure.lattice.matrix, dtype=datatype, requires_grad=True
    128 )
--> 129 center_index, neighbor_index, image, distance = structure.get_neighbor_list(
    130     r=self.atom_graph_cutoff, sites=structure.sites, numerical_tol=1e-8
    131 )
    133 # Make Graph
    134 graph = self.create_graph(
    135     n_atoms, center_index, neighbor_index, image, distance
    136 )


File ~\anaconda3\lib\site-packages\pymatgen\core\structure.py:1782, in IStructure.get_neighbor_list(self, r, sites, numerical_tol, exclude_self)
   1780 lattice_matrix = np.ascontiguousarray(self.lattice.matrix, dtype=float)
   1781 pbc = np.ascontiguousarray(self.pbc, dtype=int)
-> 1782 center_indices, points_indices, images, distances = find_points_in_spheres(
   1783     cart_coords,
   1784     site_coords,
   1785     r=r,
   1786     pbc=pbc,
   1787     lattice=lattice_matrix,
   1788     tol=numerical_tol,
   1789 )
   1790 cond = np.array([True] * len(center_indices))
   1791 if exclude_self:


File src\\pymatgen\\optimization\\neighbors.pyx:48, in pymatgen.optimization.neighbors.find_points_in_spheres()

ValueError: Buffer dtype mismatch, expected 'const int64_t' but got 'long'

Proposed Solution

implement the old version of chgnet or update these turorials?

Alternatives

No response

Code of Conduct

• I agree to follow this project's Code of Conduct

[BowenD-UCB]

The colab example runs smoothly on my side.
It's with pymatgen-2024.8.9 and chgnet-0.3.8

[janosh]

this is a numpy v2 compatibility issue that's specific to Windows. see materialsproject/pymatgen#3992. chgnet needs to update to support numpy v2 on Windows. in the meantime you can pip install 'numpy<2'

[J-You]

Thanks for your response. Got it!