Trainer

Source: train/Trainer.luau

Tiny training loop engine. Batches are built as real leading-dim tensors so a whole minibatch is ONE forward, ONE backward, ONE optimizer step (single autograd graph — the efficient path now that backward() accumulates per-call). t = Trainer.new({ model, optimizer, schedule?, lossFn, data, epochs, batchSize?, evalFn?, evalEvery?, onEpochEnd? }) t:fit() -> { lossHistory: {number}, evals: {number} } lossFn(model, xs, ys) -> 0-dim scalar Tensor ready for backward() - xs: batched input tensor, ys: targets array {number} Trainer.batchTensor(rows, featureShape) builds the input tensor.

Methods

batchTensor(rows: { { number } }, featureShape: { number }) -> any

Build a single tensor from rows of equal-length numeric arrays.

new(cfg: any) -> any

fit() -> any

evaluate(batchSize: number?) -> (number, number?)

Run a gradient-free evaluation: total loss over data + evalFn value.