import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader
from unlearning.Strategy import Strategy
from .wf.WF_Net import WF_Net

class WeightFiltration(Strategy):
    """
    Verbatim implementation of Poppi et al.'s WF-Net framework.
    Directly filters the convolutional weights of a target layer using a learnable
    channel mask, optimizing it via weight-space regularization.
    """
    def __init__(self, target_class_index: int, epochs: int = 10, lr: float = 0.2, gamma: float = 10.0):
        super().__init__(target_class_index=target_class_index)
        self.epochs = epochs
        self.lr = lr
        self.gamma = gamma 
        #self.alpha = None





    def _optimise_filter(self, model: nn.Module, forget_loader: DataLoader, retain_loader: DataLoader, device):
        # 1. Initialize the wrapper with your pre-trained model
        num_classes = model.fc.out_features
        wf_model = WF_Net(original_model=model, num_classes=num_classes).to(device)
        
        # 2. ONLY optimize alpha (everything else is frozen inside the wrapper)
        optimizer = optim.Adam([wf_model.alpha], lr=self.lr)
        criterion = nn.CrossEntropyLoss()

        for epoch in range(self.epochs):
            forget_iter = iter(forget_loader)
            t_loss_r, t_loss_f = 0.0, 0.0
            steps = 0
            
            for r_inputs, r_labels in retain_loader:
                r_inputs, r_labels = r_inputs.to(device), r_labels.to(device)
                
                # Pull the matching forget batch input
                try:
                    f_inputs, _ = next(forget_iter)
                except StopIteration:
                    forget_iter = iter(forget_loader)
                    f_inputs, _ = next(forget_iter)
                f_inputs = f_inputs.to(device)
                
                optimizer.zero_grad()
                
                # --- APPLY ALGORITHM 1 FORWARD PASS TO BOTH INPUTS ---
                # Pass the input batch AND the target unlearn class index
                outputs_r = wf_model(r_inputs, target_unlearn_class=self.target_class_index)
                outputs_f = wf_model(f_inputs, target_unlearn_class=self.target_class_index)
                
                # Compute Losses using Poppi et al.'s temperature scaled entropy
                loss_r = criterion(outputs_r, r_labels)
                
                temperature = 3.0
                logits_f_scaled = outputs_f / temperature
                
                # Compute uniform target entropy per-sample, then average over the batch
                log_probs_f = torch.log_softmax(logits_f_scaled, dim=-1)
                uniform_target = torch.ones_like(logits_f_scaled) / num_classes
                loss_f = -torch.sum(uniform_target * log_probs_f, dim=-1).mean()
                                
                total_loss = loss_r + (self.gamma * loss_f)
                total_loss.backward()
                optimizer.step()
                                
                t_loss_r += loss_r.item()
                t_loss_f += loss_f.item()
                steps += 1
                            
            print(f"  Epoch {epoch+1}/{self.epochs} | Retain Loss: {t_loss_r/steps:.4f} | Forget Loss: {t_loss_f/steps:.4f}")
        return wf_model
            


    def _run(self, model: nn.Module, forget_loader: DataLoader, retain_loader: DataLoader) -> nn.Module:
        device = next(model.parameters()).device
        model.eval() 
        
        # In WF-Net, the mask targets the last major convolutional block
        # For ResNet-18, that is the final conv layer in layer4 block 1
        if hasattr(model, 'layer4') and len(model.layer4) > 1:
            target_conv = model.layer4[1].conv2
        else:
            raise AttributeError("Model architecture does not match expected ResNet-18 structure.")
            
        # Store a pristine, non-grad copy of the original trained weights
        # Shape of conv2.weight: (out_channels, in_channels, kernel_size, kernel_size) -> e.g., (512, 512, 3, 3)
        original_weights = target_conv.weight.data.clone().detach()
        out_channels = original_weights.shape[0]

        # Initialize alpha gate vector matching Poppi et al.'s initialization range
        # Shape: (out_channels,) -> acting directly as a filter-level gate
        #self.alpha = nn.Parameter(torch.ones(out_channels, device=device) * 1.5)

        # Freeze the global model graph; only optimize our filter parameter mask
        for p in model.parameters():
            p.requires_grad = False
        #self.alpha.requires_grad = True

        wf_model = self._optimise_filter(
            model,
            forget_loader=forget_loader,
            retain_loader=retain_loader,
            device=device,
        )

        # --- PERMANENT BAKING STEP ---
        with torch.no_grad():
            # Grab the alpha mask vector for the forgotten class and cast to 4D tensor shape
            final_mask = torch.sigmoid(wf_model.alpha[self.target_class_index]).view(-1, 1, 1, 1)
            
            # Apply filter masking permanently back onto the base layer
            target_conv.weight.copy_(original_weights * final_mask)
                        
            # Unfreeze architecture parameters for evaluations downstream
            for p in model.parameters():
                p.requires_grad = True
                        
            print(f">> Permanently altered {out_channels} convolutional filters in layer4 via WF-Net.")
        return model