LAP/torchlap/modules/adaptive_lap.py

from typing import List, Type

from torch import nn
import torch

from .lap import LAP, LAPRelProp
from ..interpreting.relcam.relprop import RPProvider


class AdaptiveUnfold(nn.Module):

    def __init__(self):
        super().__init__()

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        return x.flatten(start_dim=1).unsqueeze(-1)


class AdaptiveLAP(LAP):

    def __init__(self, in_channels: int, hidden_channels: List[int] = [],
                 sigmoid_scale: float = 1.0,
                 n_attention_heads: int = 1,
                 hidden_activation: Type[nn.Module] = nn.ReLU,
                 discriminative_attention: bool = False) -> None:
        super().__init__(in_channels,
                         hidden_channels=hidden_channels,
                         sigmoid_scale=sigmoid_scale,
                         n_attention_heads=n_attention_heads,
                         hidden_activation=hidden_activation,
                         discriminative_attention=discriminative_attention)

        self.unfold = AdaptiveUnfold()

    def _get_Hp(self, _: int) -> int:
        return 1

    def _get_Wp(self, _: int) -> int:
        return 1


@RPProvider.register(AdaptiveLAP)
class AdaptiveLAPRelProp(LAPRelProp):

    def fold(self, w: torch.Tensor, size: torch.Size) -> torch.Tensor:
        return w.reshape(-1, 1, *size)