Function tensor_accuracy 

pub fn tensor_accuracy<B>(
    logits: &Tensor<B>,
    targets: &Tensor<B>,
) -> Result<f64, Error>
where
    B: Backend,

Compute accuracy directly from logit tensors and one-hot/class-index targets.

• If targets has shape [batch, n_classes] (one-hot), takes argmax of both.
• If targets has shape [batch] or [batch, 1], treats as class indices.