model_utils.py

import torch
import torch.nn as nn
from transformers import Wav2Vec2Model

class ForensicDeepfakeDetector(nn.Module):
    def __init__(self, model_name="facebook/wav2vec2-xls-r-300m"):
        super().__init__()
        # 1. Multilingual Backbone (Arabic/English)
        self.backbone = Wav2Vec2Model.from_pretrained(model_name)
        
        # Freeze to fit 8GB VRAM
        for param in self.backbone.parameters():
            param.requires_grad = False
            
        # 2. Adaptive Pooling (Handles variable lengths)
        self.pooling = nn.AdaptiveAvgPool1d(1)
        
        # 3. Classifier (Outputs raw Logits)
        self.classifier = nn.Sequential(
            nn.Linear(1024, 512),
            nn.LeakyReLU(0.2),
            nn.Dropout(0.3),
            nn.Linear(512, 1) # No Sigmoid here! Stability is handled by the Loss function
        )

    def forward(self, x):
        with torch.no_grad():
            outputs = self.backbone(x).last_hidden_state
        
        outputs = outputs.permute(0, 2, 1) # [B, 1024, Time]
        pooled = self.pooling(outputs).squeeze(-1) # [B, 1024]
        return self.classifier(pooled)