import torch
import torch.nn as nn
import torch.nn.functional as F

class WF_Net(nn.Module):
    """
    Implements Poppi et al.'s WF Model structure.
    Wraps a pre-trained ResNet-18 and dynamically applies 
    weight-space gating matrix multiplication during the forward step.
    """
    def __init__(self, original_model: nn.Module, num_classes: int):
        super().__init__()
        # Extract the sequence of blocks/layers L from the original model
        self.conv1 = original_model.conv1
        self.bn1 = original_model.bn1
        self.relu = original_model.relu
        self.maxpool = original_model.maxpool
        
        self.layer1 = original_model.layer1
        self.layer2 = original_model.layer2
        self.layer3 = original_model.layer3
        self.layer4 = original_model.layer4
        
        self.avgpool = original_model.avgpool
        self.fc = original_model.fc
        
        # Target layer for filtering: layer4 block 1 conv2
        # We extract its static tensor data out of the autograd parameter pool
        self.target_conv = self.layer4[1].conv2
        self.original_w = nn.Parameter(self.target_conv.weight.data.clone().detach(), requires_grad=False)
        
        # Require: Alpha gating matrix. Shape: (num_classes, out_channels)
        # Initialized to 1.5 as per Poppi et al.'s verbatim specification
        out_channels = self.original_w.shape[0]
        #self.alpha = nn.Parameter(torch.ones(num_classes, out_channels) * 1.5)
        self.alpha = nn.Parameter(torch.ones(num_classes, out_channels))

    def forward(self, x: torch.Tensor, target_unlearn_class: int) -> torch.Tensor:
        """
        Implements Algorithm 1: General forward step of a WF model
        Inputs: 
          x: Input tensor (Xin)
          target_unlearn_class: The class index we are actively filtering out (Yunl)
        """
        # 1. Run through early sequence of layers undisturbed
        x = self.maxpool(self.relu(self.bn1(self.conv1(x))))
        x = self.layer1(x)
        x = self.layer2(x)
        x = self.layer3(x)
        
        # Run layer4 block 0 and block 1 conv1 normally
        x = self.layer4[0](x)

        identity = x



        x = self.layer4[1].conv1(x)
        x = self.layer4[1].bn1(x)
        x = self.layer4[1].relu(x)
        
        # 2. CORE WF-NET MATH: w_hat_l <- alpha_l[Yunl] ⊙ w_l
        # Extract 1D vector for target class and reshape to (out_channels, 1, 1, 1) for 4D convolution broadcasting
        mask = torch.sigmoid(self.alpha[target_unlearn_class]).view(-1, 1, 1, 1)
        w_hat = self.original_w * mask
        
        # 3. Pass gated weights straight to functional forward pass: l(Xi, w_hat_l)
        x = F.conv2d(
            x, 
            weight=w_hat, 
            bias=self.target_conv.bias, 
            stride=self.target_conv.stride, 
            padding=self.target_conv.padding
        )
        x = self.layer4[1].bn2(x)
        
        # Handle residual shortcut skip connection manually since we opened up block 1
        # In ResNet-18 layer4, block 1 has no downsample shortcut layer; it's a direct identity add
        x = self.layer4[1].relu(x + identity) 
        
        # 4. Final Classification Head Sequence
        x = self.avgpool(x)
        x = torch.flatten(x, 1)
        y_out = self.fc(x)
        
        return y_out