/onnx/reference/ops/op_negative_log_likelihood_loss.py

# Copyright (c) ONNX Project Contributors

# SPDX-License-Identifier: Apache-2.0
from __future__ import annotations

import numpy as np

from onnx.reference.op_run import OpRun


def _compute_negative_log_likelihood_loss(
    x, target, weight=None, reduction="mean", ignore_index=None
):
    input_shape = x.shape
    if len(input_shape) == 1:
        raise RuntimeError(f"Unsupported shape {input_shape!r}.")

    target_shape = target.shape
    N = input_shape[0]
    C = input_shape[1]

    # initialize the positional weights when required
    gather_weight = None
    if weight is not None:
        # setting mode='clip' to deal with ignore_index > C or < 0 cases.
        # when the target value is > C or < 0, it doesn't matter which value we are
        # taking in gather_weight, since it will be set to 0 in the following if-block
        # use np.int32 to make it compatible with x86 machines
        gather_weight = np.take(weight, np.array(target, dtype=np.int32), mode="clip")
        # set `ignore_index`'s loss weight to 0.
        # The loss tensor will be multiplied by this weight tensor,
        # so `ignore_index`'s loss value will be eliminated.
        if ignore_index is not None:
            gather_weight = np.where(target == ignore_index, 0, gather_weight).astype(
                dtype=x.dtype
            )
    elif ignore_index != -1:
        gather_weight = np.where(target == ignore_index, 0, 1).astype(dtype=x.dtype)

    # if input is 4-d and above, make it 3-d
    if len(input_shape) != 3:
        x = x.reshape((N, C, -1))
        target = target.reshape((N, -1))

    # Get a dimension from the reshaped input.
    # If the original input shape is [N, C, H, W],
    # the D here should be H * W because we reshape
    # [N, C, H, W] to [N, C, H * W].
    D = x.shape[2]
    neg_gather_element_input = np.zeros((N, D), dtype=x.dtype)
    for i in range(N):
        for d in range(D):
            if target[i][d] != ignore_index:
                neg_gather_element_input[i][d] = -x[i][target[i][d]][d]

    loss = neg_gather_element_input

    # if the input was 4-d or above reshape to the right shape
    if len(input_shape) != 3:
        loss = loss.reshape(target_shape)

    # apply the weights when required
    if gather_weight is not None:
        loss = gather_weight * loss
        if reduction == "mean":
            loss = loss.sum() / gather_weight.sum()
            return (loss,)

    if reduction == "mean":
        loss = np.mean(loss)
    elif reduction == "sum":
        loss = np.sum(loss)
    return (loss.astype(x.dtype),)


class NegativeLogLikelihoodLoss(OpRun):
    def _run(self, x, target, weight=None, ignore_index=None, reduction=None):
        return _compute_negative_log_likelihood_loss(
            x,
            target,
            weight=weight,
            reduction=reduction,
            ignore_index=ignore_index,
        )