pypto-case-loss-crossentropy

Cross-entropy loss is computed by implementing a multi-stage Pypto kernel that performs per-sample softmax and gather operations. This is followed by a batch-wide reduction to output a final scalar tensor representing the averaged loss.

Two-stage tiling in Pypto splits cross-entropy loss computation into two explicit tile configurations. The first phase handles per-sample softmax and gather, while the second phase performs the batch reduction to generate the final scalar output.

Yes, Pypto supports decorator-based kernel definitions to compute cross-entropy loss. You can instantiate the kernel with batch size and class count, run it on predictions and targets, and obtain a scalar tensor output for training.

Pypto classification kernels handle integer targets through explicit type-casting between INT64 targets and INT32 indices. The kernel performs index-based gathering after per-sample softmax to extract the relevant prediction scores before batch reduction.

Pypto supports cross-entropy loss for batched classification by processing multiple input predictions alongside integer targets. The kernel applies two-stage tile configurations to manage the per-sample operations and final batch-wide reduction efficiently.

Multi-stage tiling for cross-entropy loss is appropriate when building Pypto-based kernels that require explicit type annotations and decorator usage. It structurally separates per-sample softmax and gather operations from the final batch-wide reduction.