Module neural_compute 

Auto-generated module structure

Modules

feedforwardnet_traits

FeedForwardNet - Trait Implementations

inferencepipeline_traits

InferencePipeline - Trait Implementations

messagepassingnet_traits

MessagePassingNet - Trait Implementations

Structs

AdamOptimizer

Adam optimizer for a flat parameter vector.

AtomicNeuralNetwork

Atomic neural network potential (NNP) with one sub-network per element.

AttentionReadout

Attention-based graph readout that computes a weighted sum of node features.

BatchNormLayer

Batch normalization layer (inference mode).

BehlerParrinelloDescriptor

Behler-Parrinello symmetry functions for constructing atomic descriptors.

Conv1DLayer

A 1-D convolutional layer operating on a sequence of feature vectors.

DataNormalizer

Feature-wise Z-score normalizer.

DenseLayer

A single fully-connected (dense) layer with an activation function.

DenseLayer64

A fully-connected layer with f64 weights supporting forward pass and gradient computation for backpropagation.

DropoutLayer

Dropout regularisation layer.

FeedForwardNet

A sequential feed-forward neural network.

GnnLayer

A single graph neural network layer implementing the sum-aggregation message passing update:

GpuNeuralBuffer

A mock GPU buffer for batched neural network inference.

GradAccumulator

Accumulates gradients from multiple backward passes for mini-batch training.

InferencePipeline

An inference pipeline that chains DenseLayer and BatchNormLayer operations.

LayerNorm

Layer normalisation applied to each time step independently.

LayerNormLayer

Layer normalisation (Ba et al., 2016).

MessagePassingNet

A multi-layer message passing neural network stacking GnnLayers.

MultiHeadAttention

Multi-head attention module.

NetworkBuilder

Convenience builder for standard AANN architectures.

NeuralLayer

A single fully-connected layer with f64 weights.

NeuralNetwork

A sequential feed-forward neural network using f64 precision.

PositionalEncoding

Sinusoidal positional encoding (Vaswani et al., 2017).

RnnCell

A single-step Elman RNN cell: h_t = activation(W_x * x_t + W_h * h_{t-1} + b).

TransformerBlock

A single transformer encoder block: x → MHA(LayerNorm(x)) + x → FFN(LayerNorm(·)) + ·

TransformerFfn

Position-wise feed-forward network used inside a transformer block.

Enums

ActivationFn

Activation functions for neural network layers.

ActivationFn64

Activation function for f64-precision neural network layers.

ExtActivation

Extended activation functions with additional variants for f64 paths.

InferenceOp

A single operation in the inference pipeline.

Functions

batch_atomic_energies

Predict atomic energies for a batch of (descriptor, atomic_number) pairs.

batch_forward

Run a forward pass for every sample in batch.

clip_gradients_by_norm

Clip a collection of gradient slices in-place so that their combined L2 norm does not exceed max_norm.

compute_forces_batch

Run the network on every position and return a force vector per atom.

compute_gradient_norm

Compute the L2 norm (Frobenius norm) of a concatenated gradient vector.

cross_entropy_loss

Compute cross-entropy loss between predictions and one-hot target.

huber_loss

Huber loss between prediction and target.

huber_loss_grad

Gradient of Huber loss w.r.t. predictions.

l2_regularisation

Compute L2 regularisation loss contribution.

l2_regularisation_grad

Compute L2 regularisation gradient contribution (adds lambda * w to each element).

load_weights_from_buffer

Load weights from a flat f32 buffer into a network, partitioning by layer sizes.

mean_huber_loss

Mean Huber loss over a batch.

mse_loss

Mean squared error loss between predictions and targets.

neural_potential_energy

Compute a scalar potential energy by summing the first network output over all atoms.

save_weights_to_buffer

Serialize network weights to a flat f32 buffer.

scaled_dot_product_attention

Compute scaled dot-product attention.

softmax

Compute softmax of a vector.