"""Small scratch-only CNN used for both DeepWeeds comparison runs."""

from __future__ import annotations

from torch import Tensor, nn


class SmallDeepWeedsCNN(nn.Module):
    """Three-block CNN with no pretrained parameters or regularization layers."""

    provenance = "scratch"

    def __init__(self, num_classes: int = 9, in_channels: int = 3) -> None:
        super().__init__()
        if num_classes <= 0:
            raise ValueError("num_classes must be positive")
        if in_channels not in (1, 3):
            raise ValueError("in_channels must be 1 (grayscale) or 3 (RGB)")

        self.features = nn.Sequential(
            nn.Conv2d(in_channels, 16, kernel_size=3, padding=1),
            nn.ReLU(),
            nn.MaxPool2d(kernel_size=2),
            nn.Conv2d(16, 32, kernel_size=3, padding=1),
            nn.ReLU(),
            nn.MaxPool2d(kernel_size=2),
            nn.Conv2d(32, 64, kernel_size=3, padding=1),
            nn.ReLU(),
            nn.AdaptiveAvgPool2d(output_size=1),
        )
        self.classifier = nn.Sequential(
            nn.Flatten(),
            nn.Linear(64, num_classes),
        )

    def forward(self, inputs: Tensor) -> Tensor:
        """Return unnormalized class logits."""

        return self.classifier(self.features(inputs))