Merge branch 'main' of https://github.com/materialsvirtuallab/matgl i…

…nto tko

changes.md

@@ -6,6 +6,10 @@ nav_order: 3
 
 # Change Log
 
+## 0.8.1
+- Fixed bug with loading of models trained with GPUs.
+- Updated default model for relaxations to be the `M3GNet-MP-2021.2.8-DIRECT-PES model`.
+
 ## 0.8.0
 - Fix a bug with use of set2set in M3Gnet implementation that affected intensive models such as the formation energy
   model. M3GNet model version is updated to 2 to invalidate previous models. Note that PES models are unaffected.

docs/changes.md

@@ -6,6 +6,18 @@ nav_order: 3
 
 # Change Log
 
+## 0.8.1
+- Fixed bug with loading of models trained with GPUs.
+- Updated default model for relaxations to be the `M3GNet-MP-2021.2.8-DIRECT-PES model`.
+
+## 0.8.0
+- Fix a bug with use of set2set in M3Gnet implementation that affected intensive models such as the formation energy
+  model. M3GNet model version is updated to 2 to invalidate previous models. Note that PES models are unaffected.
+  (@kenko911)
+
+## 0.7.1
+- Minor optimizations for memory and isolated atom training (@kenko911)
+
 ## 0.7.0
 - MatGL now supports structures with isolated atoms. (@JiQi535)
 - Fourier expansion layer and generalize cutoff polynomial. (@lbluque)

docs/index.md

@@ -8,7 +8,7 @@ nav_order: 1
 [![Linting](https://github.com/materialsvirtuallab/matgl/workflows/Linting/badge.svg)](https://github.com/materialsvirtuallab/matgl/workflows/Linting/badge.svg)
 [![Testing](https://github.com/materialsvirtuallab/matgl/workflows/Testing/badge.svg)](https://github.com/materialsvirtuallab/matgl/workflows/Testing/badge.svg)
 [![Downloads](https://pepy.tech/badge/matgl)](https://pepy.tech/project/matgl)
-[![Coverage Status](https://coveralls.io/repos/github/materialsvirtuallab/matgl/badge.svg?branch=main)](https://coveralls.io/github/materialsvirtuallab/matgl?branch=main)
+[![codecov](https://codecov.io/gh/materialsvirtuallab/matgl/branch/main/graph/badge.svg?token=3V3O79GODQ)](https://codecov.io/gh/materialsvirtuallab/matgl)
 
 # Materials Graph Library <img src="https://github.com/materialsvirtuallab/matgl/blob/main/assets/MatGL.png?raw=true" alt="matgl" width="30%" style="float: right">
 {: .no_toc }

docs/matgl.ext.md

@@ -106,7 +106,7 @@ Perform calculation for an input Atoms.
 Properties calculator can handle (energy, forces, …)
 
 
-### _class_ matgl.ext.ase.MolecularDynamics(atoms: Atoms, potential: [Potential](matgl.apps.md#matgl.apps.pes.Potential), state_attr: torch.Tensor | None = None, ensemble: str = 'nvt', temperature: int = 300, timestep: float = 1.0, pressure: float = 6.324209121801212e-07, taut: float | None = None, taup: float | None = None, compressibility_au: float | None = None, trajectory: str | Trajectory | None = None, logfile: str | None = None, loginterval: int = 1, append_trajectory: bool = False)
+### _class_ matgl.ext.ase.MolecularDynamics(atoms: Atoms, potential: [Potential](matgl.apps.md#matgl.apps.pes.Potential), state_attr: torch.Tensor | None = None, ensemble: Literal['nvt', 'npt', 'npt_berendsen'] = 'nvt', temperature: int = 300, timestep: float = 1.0, pressure: float = 6.324209121801212e-07, taut: float | None = None, taup: float | None = None, compressibility_au: float | None = None, trajectory: str | Trajectory | None = None, logfile: str | None = None, loginterval: int = 1, append_trajectory: bool = False)
 Bases: `object`
 
 Molecular dynamics class.
@@ -250,7 +250,7 @@ Relax an input Atoms.
 
 
 ### _class_ matgl.ext.ase.TrajectoryObserver(atoms: Atoms)
-Bases: `object`
+Bases: `Sequence`
 
 Trajectory observer is a hook in the relaxation process that saves the
 intermediate structures.
@@ -264,8 +264,8 @@ Init the Trajectory Observer from a Atoms.
 
 
 
-#### compute_energy()
-Calculate the potential energy.
+#### as_pandas()
+Returns: DataFrame of energies, forces, streeses, cells and atom_positions.
 
 
 #### save(filename: str)

docs/matgl.graph.md

@@ -51,6 +51,32 @@ bond_vec (torch.tensor): bond distance between two atoms
 bond_dist (torch.tensor): vector from src node to dst node
 
 
+### matgl.graph.compute.compute_theta(edges: EdgeBatch, cosine: bool = False)
+User defined dgl function to calculate bond angles from edges in a graph.
+
+
+* **Parameters**
+
+
+    * **edges** – DGL graph edges
+
+
+    * **cosine** – Whether to return the cosine of the angle or the angle itself
+
+
+
+* **Returns**
+
+    Dictionary containing bond angles and distances
+
+
+
+* **Return type**
+
+    dict[str, torch.Tensor]
+
+
+
 ### matgl.graph.compute.compute_theta_and_phi(edges: EdgeBatch)
 Calculate bond angle Theta and Phi using dgl graphs.
 
@@ -63,14 +89,14 @@ phi: torch.Tensor
 triple_bond_lengths (torch.tensor):
 
 
-### matgl.graph.compute.create_line_graph(g_batched: DGLGraph, threebody_cutoff: float)
+### matgl.graph.compute.create_line_graph(g: DGLGraph, threebody_cutoff: float)
 Calculate the three body indices from pair atom indices.
 
 
 * **Parameters**
 
 
-    * **g_batched** – Batched DGL graph
+    * **g** – DGL graph
 
 
     * **threebody_cutoff** (*float*) – cutoff for three-body interactions
@@ -107,7 +133,7 @@ Returns:
 DGLGraph object, state_attr
 
 
-#### get_graph_from_processed_structure(structure, src_id, dst_id, images, lattice_matrix, Z, element_types, cart_coords)
+#### get_graph_from_processed_structure(structure, src_id, dst_id, images, lattice_matrix, element_types, cart_coords)
 Construct a dgl graph from processed structure and bond information.
 
 
@@ -129,9 +155,6 @@ Construct a dgl graph from processed structure and bond information.
     * **lattice_matrix** – lattice information of the structure.
 
 
-    * **Z** – Atomic number information of all atoms in the structure.
-
-
     * **element_types** – Element symbols of all atoms in the structure.
 
 
@@ -149,7 +172,7 @@ Construct a dgl graph from processed structure and bond information.
 Tools to construct a dataset of DGL graphs.
 
 
-### _class_ matgl.graph.data.M3GNetDataset(structures: list, energies: list, forces: list, stresses: None | list, converter: GraphConverter, threebody_cutoff: float, name='M3GNETDataset', graph_labels: list | None = None)
+### _class_ matgl.graph.data.M3GNetDataset(converter: GraphConverter, threebody_cutoff: float, structures: list, energies: list | None = None, forces: list | None = None, stresses: list | None = None, labels: list | None = None, name='M3GNETDataset', label_name: str | None = None, graph_labels: list | None = None)
 Bases: `DGLDataset`
 
 Create a dataset including dgl graphs.
@@ -158,6 +181,12 @@ Create a dataset including dgl graphs.
 * **Parameters**
 
 
+    * **converter** – dgl graph converter
+
+
+    * **threebody_cutoff** – cutoff for three body
+
+
     * **structures** – Pymatgen structure
 
 
@@ -170,13 +199,13 @@ Create a dataset including dgl graphs.
     * **stresses** – Target stresses
 
 
-    * **converter** – dgl graph converter
+    * **labels** – target properties
 
 
-    * **threebody_cutoff** – cutoff for three body
+    * **name** – name of dataset
 
 
-    * **name** – name of dataset
+    * **label_name** – name of target properties
 
 
     * **graph_labels** – state attributes.
@@ -333,7 +362,7 @@ Dataloader for MEGNet training.
 
 
 
-### matgl.graph.data.collate_fn(batch)
+### matgl.graph.data.collate_fn(batch, include_line_graph: bool = False)
 Merge a list of dgl graphs to form a batch.
 
 

docs/matgl.layers.md

@@ -249,7 +249,7 @@ increasing number of roots and a given cutoff.
 
 Details are given in: [https://arxiv.org/abs/2003.03123](https://arxiv.org/abs/2003.03123)
 
-This is equivalent to SphericalBesselFunction class with max_l=1, i.e. only l=0 bessel fucntions), but with
+This is equivalent to SphericalBesselFunction class with max_l=1, i.e. only l=0 bessel functions), but with
 optional learnable frequencies.
 
 
@@ -1061,7 +1061,7 @@ Perform attribute (global state) update.
 Readout layer for M3GNet.
 
 
-### _class_ matgl.layers._readout.ReduceReadOut(op: str = 'mean', field: str = 'node_feat')
+### _class_ matgl.layers._readout.ReduceReadOut(op: str = 'mean', field: Literal['node_feat', 'edge_feat'] = 'node_feat')
 Bases: `Module`
 
 Reduce atom or bond attributes into lower dimensional tensors as readout.
@@ -1094,7 +1094,7 @@ This could be summing up the atoms or bonds, or taking the mean, etc.
 
 #### training(_: boo_ )
 
-### _class_ matgl.layers._readout.Set2SetReadOut(num_steps: int, num_layers: int, field: str)
+### _class_ matgl.layers._readout.Set2SetReadOut(in_feats: int, n_iters: int, n_layers: int, field: Literal['node_feat', 'edge_feat'])
 Bases: `Module`
 
 The Set2Set readout function.
@@ -1103,10 +1103,13 @@ The Set2Set readout function.
 * **Parameters**
 
 
-    * **num_steps** (*int*) – Number of LSTM steps
+    * **in_feats** (*int*) – length of input feature vector
 
 
-    * **num_layers** (*int*) – Number of layers.
+    * **n_iters** (*int*) – Number of LSTM steps
+
+
+    * **n_layers** (*int*) – Number of layers.
 
 
     * **field** (*str*) – Field of graph to perform the readout.

docs/matgl.md

@@ -107,7 +107,7 @@ MatGL (Materials Graph Library) is a graph deep learning library for materials s
         * [`TrajectoryObserver`](matgl.ext.md#matgl.ext.ase.TrajectoryObserver)
 
 
-            * [`TrajectoryObserver.compute_energy()`](matgl.ext.md#matgl.ext.ase.TrajectoryObserver.compute_energy)
+            * [`TrajectoryObserver.as_pandas()`](matgl.ext.md#matgl.ext.ase.TrajectoryObserver.as_pandas)
 
 
             * [`TrajectoryObserver.save()`](matgl.ext.md#matgl.ext.ase.TrajectoryObserver.save)
@@ -145,6 +145,9 @@ MatGL (Materials Graph Library) is a graph deep learning library for materials s
         * [`compute_pair_vector_and_distance()`](matgl.graph.md#matgl.graph.compute.compute_pair_vector_and_distance)
 
 
+        * [`compute_theta()`](matgl.graph.md#matgl.graph.compute.compute_theta)
+
+
         * [`compute_theta_and_phi()`](matgl.graph.md#matgl.graph.compute.compute_theta_and_phi)
 
 

docs/matgl.models.md

@@ -23,7 +23,7 @@ Computational Science, 2023, 2, 718-728. DOI: 10.1038/s43588-022-00349-3.
 ```
 
 
-### _class_ matgl.models._m3gnet.M3GNet(element_types: tuple[str], dim_node_embedding: int = 64, dim_edge_embedding: int = 64, dim_state_embedding: int | None = None, dim_state_types: int | None = None, dim_state_feats: int | None = None, max_n: int = 3, max_l: int = 3, nblocks: int = 3, rbf_type='SphericalBessel', is_intensive: bool = True, readout_type: str = 'weighted_atom', task_type: str = 'regression', cutoff: float = 5.0, threebody_cutoff: float = 4.0, units: int = 64, ntargets: int = 1, use_smooth: bool = False, use_phi: bool = False, niters_set2set: int = 3, nlayers_set2set: int = 3, field: str = 'node_feat', include_state: bool = False, activation_type: str = 'swish', \*\*kwargs)
+### _class_ matgl.models._m3gnet.M3GNet(element_types: tuple[str], dim_node_embedding: int = 64, dim_edge_embedding: int = 64, dim_state_embedding: int | None = None, ntypes_state: int | None = None, dim_state_feats: int | None = None, max_n: int = 3, max_l: int = 3, nblocks: int = 3, rbf_type='SphericalBessel', is_intensive: bool = True, readout_type: str = 'weighted_atom', task_type: str = 'regression', cutoff: float = 5.0, threebody_cutoff: float = 4.0, units: int = 64, ntargets: int = 1, use_smooth: bool = False, use_phi: bool = False, niters_set2set: int = 3, nlayers_set2set: int = 3, field: Literal['node_feat', 'edge_feat'] = 'node_feat', include_state: bool = False, activation_type: str = 'swish', \*\*kwargs)
 Bases: `Module`, [`IOMixIn`](matgl.utils.md#matgl.utils.io.IOMixIn)
 
 The main M3GNet model.
@@ -47,7 +47,7 @@ The main M3GNet model.
     * **dim_state_feats** (*int*) – number of state features after linear layer
 
 
-    * **dim_state_types** (*int*) – number of state labels
+    * **ntypes_state** (*int*) – number of state labels
 
 
     * **max_n** (*int*) – number of radial basis expansion

docs/tutorials/Property Predictions using MEGNet or M3GNet Models.md

@@ -49,21 +49,16 @@ eform = model.predict_structure(struct)
 print(f"The predicted formation energy for CsCl is {float(eform):.3f} eV/atom.")
 ```
 
-    The predicted formation energy for CsCl is -2.272 eV/atom.
-
-
 ## Using the M3GNet-MP-2018.6.1-Eform model
 
 
 ```python
+# Load the pre-trained M3GNet formation energy model
 model = matgl.load_model("M3GNet-MP-2018.6.1-Eform")
 eform = model.predict_structure(struct)
 print(f"The predicted formation energy for CsCl is {float(eform):.3f} eV/atom.")
 ```
 
-    The predicted formation energy for CsCl is -2.233 eV/atom.
-
-
 # MP Band gap
 
 This is the multi-fidelity band gap model, discussed in Chen, C.; Zuo, Y.; Ye, W.; Li, X.; Ong, S. P. Learning Properties of Ordered and Disordered Materials from Multi-Fidelity Data. Nature Computational Science 2021, 1, 46–53. https://doi.org/10.1038/s43588-020-00002-x.
@@ -82,9 +77,4 @@ for i, method in ((0, "PBE"), (1, "GLLB-SC"), (2, "HSE"), (3, "SCAN")):
         structure=struct, state_feats=graph_attrs
     )
     print(f"The predicted {method} band gap for CsCl is {float(bandgap):.3f} eV.")
-```
-
-    The predicted PBE band gap for CsCl is 4.825 eV.
-    The predicted GLLB-SC band gap for CsCl is 8.323 eV.
-    The predicted HSE band gap for CsCl is 6.317 eV.
-    The predicted SCAN band gap for CsCl is 5.965 eV.
+```

matgl/graph/compute.py

@@ -104,11 +104,7 @@ def compute_theta_and_phi(edges: dgl.udf.EdgeBatch):
     triple_bond_lengths (torch.tensor):
     """
     angles = compute_theta(edges, cosine=True)
-    angles.update(
-        {
-            "phi": torch.zeros_like(angles["cos_theta"]),
-        }
-    )
+    angles["phi"] = torch.zeros_like(angles["cos_theta"])
     return angles
 
 

matgl/utils/io.py

@@ -111,11 +111,9 @@ def load(cls, path: str | Path | dict, **kwargs):
 
         _check_ver(cls, model_data)
 
-        if not torch.cuda.is_available():
-            state = torch.load(fpaths["state.pt"], map_location=torch.device("cpu"))
-        else:
-            state = torch.load(fpaths["state.pt"])
-        d = torch.load(fpaths["model.pt"])
+        map_location = torch.device("cpu") if not torch.cuda.is_available() else None
+        state = torch.load(fpaths["state.pt"], map_location=map_location)
+        d = torch.load(fpaths["model.pt"], map_location=map_location)
 
         # Deserialize any args that are IOMixIn subclasses.
         for k, v in d.items():

setup.py

@@ -11,7 +11,7 @@
 
 setup(
     name="matgl",
-    version="0.8.0",
+    version="0.8.1",
     author="Tsz Wai Ko, Marcel Nassar, Ji Qi, Santiago Miret, Eliott Liu, Shyue Ping Ong",
     author_email="[email protected], [email protected]",
     maintainer="Shyue Ping Ong",