explainability.py

import os
import torch
import numpy as np
import matplotlib.pyplot as plt
from PIL import Image
from src.preprocess import val_transforms
from src.models.classifier_model import BrainHybridModel
from src.config import OUTPUT_DIR, CHECKPOINT_DIR

def generate_attention_heatmap(image_path, save_name="attention_map.png"):
    """Menghasilkan peta panas (heatmap) fokus perhatian model AI pada gambar otak"""
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    
    # 1. Muat dan ubah gambar menjadi tensor
    orig_image = Image.open(image_path).convert('RGB')
    tensor_image = val_transforms(orig_image).unsqueeze(0).to(device)
    
    # 2. Muat model dan bobot terbaik
    model = BrainHybridModel().to(device)
    checkpoint_path = os.path.join(CHECKPOINT_DIR, "best_hybrid_model.pth")
    if os.path.exists(checkpoint_path):
        try:
            model.load_state_dict(torch.load(checkpoint_path, map_location=device))
        except RuntimeError:
            print("Warning: Checkpoint tidak kompatibel dengan arsitektur ViT baru, menggunakan bobot pretrained bawaan.")
    model.eval()
    
    # 3. Ekstraksi attention weights dari ViT (custom Transformer block terakhir)
    with torch.no_grad():
        # forward_with_attention mengembalikan (logits, attn) dari block terakhir
        # attn shape: [B, num_heads, seq_len, seq_len]
        _, attentions = model.forward_with_attention(tensor_image)

        # Rata-ratakan semua attention heads
        avg_attn = attentions.squeeze(0).mean(dim=0)      # [seq_len, seq_len]

        # Ambil attention dari CLS token (index 0) ke semua patch tokens
        cls_attn = avg_attn[0, 1:]  # [num_patches] (buang CLS-to-CLS)

        # Feature map EfficientNet-B3 di 224x224 -> grid 7x7 = 49 patch tokens
        num_patches = int(cls_attn.shape[0] ** 0.5)
        heatmap = cls_attn.reshape(num_patches, num_patches).cpu().numpy()
        
        # Normalisasi peta panas antara nilai 0 hingga 1
        heatmap = np.maximum(heatmap, 0)
        heatmap /= np.max(heatmap) if np.max(heatmap) != 0 else 1.0

    # 4. Gambar dan gabungkan citra asli dengan peta panas
    fig, axes = plt.subplots(1, 2, figsize=(10, 5))
    axes[0].imshow(orig_image)
    axes[0].set_title("Gambar Medis Asli")
    axes[0].axis('off')
    
    # Ubah ukuran peta panas agar pas dengan dimensi gambar asli
    heatmap_resized = np.array(Image.fromarray(heatmap).resize(orig_image.size, Image.Resampling.BILINEAR))
    
    axes[1].imshow(orig_image)
    axes[1].imshow(heatmap_resized, cmap='jet', alpha=0.4) # Overlay warna transparan
    axes[1].set_title("Peta Fokus Atensi AI (ViT Attention)")
    axes[1].axis('off')
    
    # Simpan visualisasi ke folder outputs/figures/
    figure_dir = os.path.join(OUTPUT_DIR, "figures")
    os.makedirs(figure_dir, exist_ok=True)
    save_path = os.path.join(figure_dir, save_name)
    plt.savefig(save_path, bbox_inches='tight')
    plt.close()
    print(f"Sukses menghasilkan peta eksplanabilitas AI! Tersimpan di: {save_path}")

if __name__ == "__main__":
    # Mencari satu contoh gambar acak dari folder normal untuk uji coba modul
    sample_dir = "data/raw/Normal"
    if os.path.exists(sample_dir) and os.listdir(sample_dir):
        first_img = os.listdir(sample_dir)[0]
        full_path = os.path.join(sample_dir, first_img)
        generate_attention_heatmap(full_path)
    else:
        print("Folder data/raw/Normal kosong atau tidak ditemukan untuk pengujian.")