pypto-case-loss-crossentropy

A cross-entropy loss kernel in Pypto handles multi-input setups by applying per-sample softmax, gathering log probabilities by target indices, and reducing the batch to output a scalar loss. It uses two-phase tiling to configure different tile settings across execution stages.

Two-phase tile configuration allows different tile settings across execution stages of a cross-entropy loss kernel. You need it when running multi-input setups in Pypto to optimize per-sample softmax and batch reduction operations for scalar loss output.

You compute per-sample softmax and gather target indices by applying softmax across individual sample logits, then selecting corresponding log probabilities using target indices. The kernel sums these gathered values and performs batch reduction to compute the averaged scalar loss.

Yes, Pypto runtime supports scalar output for cross-entropy loss backpropagation. The kernel computes a single scalar loss value from batch reduction of gathered log probabilities, which can be used directly for backpropagation in your model.

Yes, you can use different tile settings for softmax and gather operations within a single cross-entropy loss kernel. Pypto's two-phase tiling mechanism allows configuring distinct tile parameters for different stages of the multi-input kernel execution.

Cross-entropy loss kernels require multi-input setups to handle both predictions and targets simultaneously during batch processing. The kernel accesses both input tensors to perform per-sample softmax on predictions and gather operations using target indices for accurate loss computation.