LAP/torchlap/modules/gaussianmax.py

import torch
from torch import nn


class GaussianMax2d(nn.Module):

    def __init__(self):
        super().__init__()
        self._eps = 1e-4
        self.alpha = nn.Parameter(torch.tensor(1.), requires_grad=True)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        r"""
        :return: exp(-alpha^2 * (x_max - x) ^ 2) * x
        """
        #   B   HW  N
        B, HW, N = x.shape
        x = x.transpose(1, 2).reshape(-1, HW)                       # BN    HW
        x_max = (x.detach() if self.training else x)\
            .max(dim=1, keepdim=True).values                        # BN    1

        out = torch.exp(-self.alpha ** 2 * (x_max - x) ** 2) * x    # BN    HW
        out = out.reshape(B, N, HW).transpose(1, 2)                 # B HW  N
        return out + self._eps