LossMaskMatrix损失函数掩码矩阵

1. 理论

$$\begin{array}{c}A=\left[\begin{array}{cccc}1& 2& 0& 0\\ & & & \\ 2& 3& 4& 0\end{array}\right]\to B=\left[\begin{array}{cccc}1& 1& 0& 0\\ & & & \\ 1& 1& 1& 0\end{array}\right]\to C=\left[\begin{array}{cccc}0.225& 0.610& 0& 0\\ & & & \\ 0.089& 0.242& 0.657& 0\end{array}\right]\end{array}$$

2. python

import torch
from torch import nn
from torch.nn import functional as F

torch.set_printoptions(precision=3, sci_mode=False)


class MatrixNoneZero2OnesLike(object):
    def __init__(self, in_matrix):
        self.in_matrix = in_matrix
        self._result = torch.zeros_like(self.in_matrix)

    @property
    def result(self):
        my_result = torch.zeros_like(self.in_matrix)
        my_result_bool = self.in_matrix.to(torch.bool)
        self._result = my_result.masked_fill(my_result_bool, 1)
        return self._result


class LossMaskedMatrix(object):
    def __init__(self, src_matrix):
        self.src_matrix = src_matrix.to(torch.float)
        self.nonzero = MatrixNoneZero2OnesLike(self.src_matrix)
        self.nonzero_matrix = self.nonzero.result
        self._loss_matrix = torch.zeros_like(self.src_matrix)

    @property
    def loss_matrix(self):
        my_soft_matrix = F.softmax(self.src_matrix, dim=-1)
        my_loss_matrix = my_soft_matrix * self.nonzero_matrix
        print(f"*" * 50)
        print(f"src_matrix=\n{self.src_matrix}")
        print(f"nonzero_matrix=\n{self.nonzero_matrix}")
        print(f"loss_matrix=\n{my_loss_matrix}")
        print(f"*" * 50)
        self._loss_matrix = my_loss_matrix

        return self._loss_matrix


if __name__ == "__main__":
    in_matrix = torch.tensor([[1, 2, 0, 0], [2, 3, 4, 0]]).to(torch.float)
    test_loss_matrix = LossMaskedMatrix(in_matrix)
    result = test_loss_matrix.loss_matrix

• 结果：

**************************************************
src_matrix=
tensor([[1., 2., 0., 0.],
        [2., 3., 4., 0.]])
nonzero_matrix=
tensor([[1., 1., 0., 0.],
        [1., 1., 1., 0.]])
loss_matrix=
tensor([[0.225, 0.610, 0.000, 0.000],
        [0.089, 0.242, 0.657, 0.000]])
**************************************************