"""Decision-theoretic point prediction for SMAPE+1.""" from __future__ import annotations import torch def smape_loss_matrix(n_classes: int, device=None, dtype=torch.float32) -> torch.Tensor: """Loss[y, c] = |y-c| / ((|y|+|c|)/2 + 1).""" y = torch.arange(int(n_classes), device=device, dtype=dtype) c = y.view(1, -1) y = y.view(-1, 1) return (y - c).abs() / ((y.abs() + c.abs()) / 2.0 + 1.0) def smape_optimal_point(probs: torch.Tensor) -> torch.Tensor: """Return argmin_c E[SMAPE+1(c)] for probs (..., C).""" if probs.dim() < 2: raise ValueError("probs must have at least 2 dimensions") C = probs.shape[-1] L = smape_loss_matrix(C, device=probs.device, dtype=probs.dtype) expected = torch.matmul(probs, L) return expected.argmin(dim=-1).to(probs.dtype) def expected_smape(probs: torch.Tensor) -> torch.Tensor: C = probs.shape[-1] L = smape_loss_matrix(C, device=probs.device, dtype=probs.dtype) expected = torch.matmul(probs, L) return expected.min(dim=-1).values def point_from_output(output: torch.Tensor, output_type: str = "regression", n_classes: int = 87) -> torch.Tensor: if output_type == "regression": return output if output_type == "distribution": probs = torch.softmax(output, dim=-1) return smape_optimal_point(probs) raise ValueError(f"Unknown output_type: {output_type}")