Module batch 

Batched Neural Evaluation

This module provides infrastructure for grouping neural predicate calls by network name, enabling efficient batched GPU evaluation.

Why Batching?

In DeepProbLog-style programs, the same neural network may be called many times:

nn(mnist_net, [X], Y, [0..9]) :: digit(X, Y).
addition(X, Y, Z) :- digit(X, LeftDigit), digit(Y, RightDigit), Z is LeftDigit + RightDigit.

For a query like addition(img1, img2, Z), we need to evaluate mnist_net twice (once for each digit). Instead of two separate forward passes, we batch them into a single mnist_net([img1, img2]) call for GPU efficiency.

Usage

use xlog_neural::batch::{BatchCollector, NeuralCall};

let mut collector = BatchCollector::new();

// During proof search, collect neural calls
collector.add(NeuralCall::new("mnist", vec![0])); // digit(img[0], Y)
collector.add(NeuralCall::new("mnist", vec![1])); // digit(img[1], Y)

// Group by network for batched evaluation
let batches = collector.collect();
let mnist_indices = collector.indices_for_network("mnist");
// mnist_indices = [0, 1] - evaluate both in one forward pass

Structs

BatchCollector

Collects neural predicate calls for batched evaluation.

BatchMapping

Mapping from call index to batch result index.

BatchResult

Result of a batched neural evaluation.

NeuralCall

A single neural predicate call.