mindchemistry.cell.orb.GraphHead

View Source On Gitee
class mindchemistry.cell.orb.GraphHead(latent_dim, num_mlp_layers, mlp_hidden_dim, target_property_dim, node_aggregation='mean', dropout=None, compute_stress=False)[source]

Graph-level prediction head. Implements graph-level prediction head that can be attached to base models for predicting graph-level properties (e.g., stress tensor) from node features using aggregation and MLP.

Parameters

  • latent_dim (int) – Input feature dimension for each node.

  • num_mlp_layers (int) – Number of hidden layers in MLP.

  • mlp_hidden_dim (int) – Hidden dimension size of MLP.

  • target_property_dim (int) – Output dimension of graph-level property.

  • node_aggregation (str, optional) – Aggregation method for node predictions, e.g., "mean" or "sum". Default: "mean".

  • dropout (Optional[float], optional) – Dropout rate for MLP. Default: None.

  • compute_stress (bool, optional) – Whether to compute and output stress tensor. Default: False.

Inputs:

  • node_features (dict) - Node feature dictionary, must contain key "feat" with shape \((n_{nodes}, latent\_dim)\).

  • n_node (Tensor) - Number of nodes in graph, shape \((1,)\).

Outputs:

  • output (dict) - Dictionary containing key "stress_pred" with value of shape \((1, target\_property\_dim)\).

Raises

ValueError – If required feature keys are missing in node_features.

Supported Platforms:

Ascend

Examples

>>> import numpy as np
>>> import mindspore
>>> from mindspore import Tensor
>>> from mindchemistry.cell.orb.gns import GraphHead
>>> graph_head = GraphHead(
...     latent_dim=256,
...     num_mlp_layers=1,
...     mlp_hidden_dim=256,
...     target_property_dim=6,
...     compute_stress=True,
... )
>>> n_atoms = 4
>>> n_node = Tensor([n_atoms], mindspore.int32)
>>> atomic_numbers = Tensor(np.random.randint(1, 119, size=(n_atoms,), dtype=np.int32))
>>> atomic_numbers_embedding_np = np.zeros((n_atoms, 118), dtype=np.float32)
>>> for i, num in enumerate(atomic_numbers.asnumpy()):
...     atomic_numbers_embedding_np[i, num - 1] = 1.0
>>> node_features = {
...     "atomic_numbers": atomic_numbers,
...     "atomic_numbers_embedding": Tensor(atomic_numbers_embedding_np),
...     "positions": Tensor(np.random.randn(n_atoms, 3).astype(np.float32)),
...     "feat": Tensor(np.random.randn(n_atoms, 256).astype(np.float32))
... }
>>> output = graph_head(node_features, n_node)
>>> print(output['stress_pred'].shape)
(1, 6)
predict(node_features, n_node, atomic_numbers=None)[source]

Predict graph-level attributes.

Parameters

  • node_features – Node features tensor

  • n_node – Number of nodes

  • atomic_numbers – Optional atomic numbers for reference energy calculation

Returns

Graph-level predictions of shape (n_graphs, target_property_dim). If compute_stress is True, this will be the stress tensor. If compute_stress is False, this will be the graph-level property (e.g., energy).

Return type

probs