xnn.common.models.ops

Tensor operations shared across model families (gnn / cnn / dnn).

Kept here, at the models level, because they are common to more than one architecture type. Per-family helpers live under the family package instead (e.g. models/gnn/base.py), and graph-level reductions that every model needs live on InteratomicPotential (aggregate_energy).

Module Attributes

ATOMIC_MASSES

Standard atomic weights (u) by atomic number, Z = 0 to 99 (the ASE table), for center_of_mass().

Functions

build_triplets(edge_index, num_nodes)	Enumerate neighbour pairs (j, k) sharing a centre i.
cell_volume(cell)	Volume of row-vector cells (..., 3, 3) as the scalar triple product.
center_of_mass(pos, atomic_numbers, batch, ...)	Center of mass of every structure of a batch.
glorot_orthogonal_(weight[, scale])	Initialize a weight matrix as a (semi-)orthogonal matrix with Glorot variance.
make_activation(activation)	Return a fresh activation module from a name or a module instance.
scatter_sum(src, index, dim_size)	Sum rows of src into dim_size buckets given by index (dim 0).
segment_sum(x, lengths)	Sum of consecutive segments of x along dim 0; lengths (int64) sum to len(x).
shifted_softplus(x)	Shifted softplus ssp(x) = ln(0.5 e^x + 0.5) with ssp(0) = 0.
softplus_inverse(x)	Return y such that softplus(y) = x (x > 0, float or array).
structure_sum(src, index, dim_size)	Sum per-atom values into per-structure totals, accumulated in float64.

Classes

GaussianActivation(*args, **kwargs)	Gaussian activation exp(-x^2) (the original ANI-1 hidden activation).
ScaledTanh(*args, **kwargs)	1.59223 tanh(x), the scaled hyperbolic tangent of the RuNNer networks.
ShiftedSoftplus(*args, **kwargs)	Module form of shifted_softplus(), for use inside nn.Sequential.
Softplus(*args, **kwargs)	Softplus log(1 + e^x) evaluated exactly at every x, as max(x, 0) + log1p(e^-|x|).
Square(*args, **kwargs)	x^2.