model.py

"""
Author: Juan Pablo Triana Martinez
LinkNet architecture — standalone for HuggingFace Spaces deployment.
"""
import torch
import torch.nn as nn


class LinknetStem(nn.Module):
    def __init__(self, m: int = 3, n: int = 64) -> None:
        super().__init__()
        self.linknet_stem = nn.Sequential(
            nn.Conv2d(m, n, kernel_size=(7, 7), stride=(2, 2), padding=(3, 3), bias=False),
            nn.BatchNorm2d(n),
            nn.ReLU(),
            nn.MaxPool2d(kernel_size=(3, 3), stride=(2, 2), padding=(1, 1)),
        )

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        return self.linknet_stem(x)


class LinknetEncoderBlock(nn.Module):
    def __init__(self, m: int, n: int) -> None:
        super().__init__()
        self.convs_blocks_1 = nn.Sequential(
            nn.Conv2d(m, n, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False),
            nn.BatchNorm2d(n),
            nn.ReLU(),
            nn.Conv2d(n, n, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False),
            nn.BatchNorm2d(n),
            nn.ReLU(),
        )
        self.skip_conn = nn.Sequential(
            nn.Conv2d(m, n, kernel_size=(1, 1), stride=(2, 2), padding=(0, 0), bias=False),
            nn.BatchNorm2d(n),
        )
        self.convs_block_2 = nn.Sequential(
            nn.Conv2d(n, n, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False),
            nn.BatchNorm2d(n),
            nn.ReLU(),
            nn.Conv2d(n, n, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False),
            nn.BatchNorm2d(n),
            nn.ReLU(),
        )

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        x1 = self.convs_blocks_1(x)
        x2 = x1 + self.skip_conn(x)
        x3 = self.convs_block_2(x2)
        return x3 + x2


class LinknetDecoderBlock(nn.Module):
    def __init__(self, m: int, n: int) -> None:
        super().__init__()
        self.conv_block_1 = nn.Sequential(
            nn.Conv2d(m, m // 4, kernel_size=(1, 1), bias=False),
            nn.BatchNorm2d(m // 4),
            nn.ReLU(),
        )
        self.upsample_block = nn.Sequential(
            nn.Upsample(scale_factor=2, mode="bilinear", align_corners=True),
            nn.Conv2d(m // 4, m // 4, kernel_size=(3, 3), padding=(1, 1), bias=False),
            nn.BatchNorm2d(m // 4),
            nn.ReLU(),
        )
        self.conv_block_2 = nn.Sequential(
            nn.Conv2d(m // 4, n, kernel_size=(1, 1), bias=False),
            nn.BatchNorm2d(n),
            nn.ReLU(),
        )

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        x = self.conv_block_1(x)
        x = self.upsample_block(x)
        return self.conv_block_2(x)


class LinknetReconstructer(nn.Module):
    def __init__(self, N: int = 1, m: int = 64, n: int = 32) -> None:
        super().__init__()
        self.upsample_block_1 = nn.Sequential(
            nn.Upsample(scale_factor=2, mode="bilinear", align_corners=True),
            nn.Conv2d(m, n, kernel_size=(3, 3), padding=(1, 1), bias=False),
            nn.BatchNorm2d(n),
            nn.ReLU(),
        )
        self.conv_block = nn.Sequential(
            nn.Conv2d(n, n, kernel_size=(3, 3), padding=(1, 1), bias=False),
            nn.BatchNorm2d(n),
            nn.ReLU(),
        )
        self.upsample_block_2 = nn.Sequential(
            nn.Upsample(scale_factor=2, mode="bilinear", align_corners=True),
            nn.Conv2d(n, N, kernel_size=(3, 3), padding=(1, 1), bias=False),
        )

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        x = self.upsample_block_1(x)
        x = self.conv_block(x)
        return self.upsample_block_2(x)


class LinknetModel(nn.Module):
    def __init__(self, Cin: int = 3, N: int = 1) -> None:
        super().__init__()
        self.stem            = LinknetStem(m=Cin, n=64)
        self.encoder_block_1 = LinknetEncoderBlock(64, 64)
        self.encoder_block_2 = LinknetEncoderBlock(64, 128)
        self.encoder_block_3 = LinknetEncoderBlock(128, 256)
        self.encoder_block_4 = LinknetEncoderBlock(256, 512)
        self.decoder_block_4 = LinknetDecoderBlock(512, 256)
        self.decoder_block_3 = LinknetDecoderBlock(256, 128)
        self.decoder_block_2 = LinknetDecoderBlock(128, 64)
        self.decoder_block_1 = LinknetDecoderBlock(64, 64)
        self.reconstructer   = LinknetReconstructer(N=N, m=64, n=32)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        x  = self.stem(x)
        x1 = self.encoder_block_1(x)
        x2 = self.encoder_block_2(x1)
        x3 = self.encoder_block_3(x2)
        x4 = self.encoder_block_4(x3)
        x  = self.decoder_block_4(x4) + x3
        x  = self.decoder_block_3(x)  + x2
        x  = self.decoder_block_2(x)  + x1
        x  = self.decoder_block_1(x)
        return self.reconstructer(x)


def create_semantic_model() -> LinknetModel:
    """Factory: LinkNet with 3-channel input and 12 semantic output channels."""
    return LinknetModel(Cin=3, N=12)