struct

`SushiAI::Data::Batch`

One minibatch, resident on the device.

Declared in: include/SushiAI/data/dataset.hpp

Public attributes

SushiBLAS::Tensor inputs

An [N, D] matrix of samples.

SushiBLAS::Tensor targets

An [N, C] one-hot target matrix.

std::vector< int32_t > labels

The class index per sample, on the host.

Kept host-side because accuracy is a metric, not training maths: scoring it needs a per-row argmax, SushiBLAS's argmax is whole-tensor, and adding a device reduction for a diagnostic would be the wrong trade.