sam/efficient_sam/mlp.py

from typing import Type

from torch import nn


# Lightly adapted from
# https://github.com/facebookresearch/MaskFormer/blob/main/mask_former/modeling/transformer/transformer_predictor.py # noqa
class MLPBlock(nn.Module):
    def __init__(
        self,
        input_dim: int,
        hidden_dim: int,
        output_dim: int,
        num_layers: int,
        act: Type[nn.Module],
    ) -> None:
        super().__init__()
        self.num_layers = num_layers
        h = [hidden_dim] * (num_layers - 1)
        self.layers = nn.ModuleList(
            nn.Sequential(nn.Linear(n, k), act())
            for n, k in zip([input_dim] + h, [hidden_dim] * num_layers)
        )
        self.fc = nn.Linear(hidden_dim, output_dim)

    def forward(self, x):
        for layer in self.layers:
            x = layer(x)
        return self.fc(x)