Struct NodeClassifier 

pub struct NodeClassifier { /* private fields */ }

A small MLP that maps NodeFeatures to a content-probability.

Implementations

impl NodeClassifier

pub fn new(device: &Device, lr: f64) -> Result<Self>

Create a new, randomly-initialized classifier.

pub fn score_batch(&self, features: &[NodeFeatures]) -> Result<Vec<f32>>

Content probability in [0, 1] for each candidate.

pub fn select_best(&self, features: &[NodeFeatures]) -> Result<Option<usize>>

Index of the candidate most likely to be the content root.

pub fn train_batch( &mut self, features: &[NodeFeatures], labels: &[f32], ) -> Result<f32>

One supervised gradient step over a batch of (features, label) pairs, where label is 1.0 for the content node and 0.0 otherwise. Returns the binary cross-entropy loss.

pub fn num_parameters(&self) -> usize

Number of trainable parameters (for logging / metadata).

pub fn save(&self, path: &Path) -> Result<()>

Serialize the classifier weights to a .safetensors file.

pub fn load(path: &Path, device: &Device, lr: f64) -> Result<Self>

Load a classifier previously written with Self::save.

The architecture is fixed (so new recreates the same variables) and the stored tensors are copied into them; lr only affects subsequent training and is irrelevant for inference.