File size: 1,402 Bytes
04ae9bb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
"""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}")