Transfer Learning: Pakai Model yang Sudah Pintar

import torch import torch.nn as nn import torchvision from torchvision import models, transforms from torch.utils.data import DataLoader # =========================== # STEP 1: Load pre-trained ResNet-18 # =========================== model = models.resnet18(weights='IMAGENET1K_V1') # ↑ Download 44.7MB model yang sudah terlatih di ImageNet # Model ini SUDAH tahu: edges, textures, shapes, objects # =========================== # STEP 2: FREEZE semua layer (jangan diubah) # =========================== for param in model.parameters(): param.requires_grad = False # 🔒 Freeze! # =========================== # STEP 3: GANTI "kepala" (FC layer terakhir) # =========================== # ResNet-18 asli: fc = Linear(512, 1000) → 1000 kelas ImageNet # Kita ganti jadi: fc = Linear(512, 10) → 10 kelas kita num_features = model.fc.in_features # 512 model.fc = nn.Linear(num_features, 10) # 🔓 Layer baru (trainable!) # Hitung: berapa parameter yang akan di-train? trainable = sum(p.numel() for p in model.parameters() if p.requires_grad) total = sum(p.numel() for p in model.parameters()) print(f"Trainable: {trainable:,} / {total:,} ({100*trainable/total:.1f}%)") # Trainable: 5,130 / 11,181,642 (0.05%) # ↑ Hanya 0.05% parameter yang di-train! Sisanya frozen. # =========================== # STEP 4: Data transforms (resize ke 224×224 untuk ResNet) # =========================== transform = transforms.Compose([ transforms.Resize(256), transforms.CenterCrop(224), # ResNet expects 224×224 transforms.ToTensor(), transforms.Normalize( [0.485, 0.456, 0.406], # ImageNet mean (RGB) [0.229, 0.224, 0.225] # ImageNet std (RGB) ) ]) # =========================== # STEP 5: Optimizer — HANYA train FC parameters! # =========================== device = torch.device("cuda" if torch.cuda.is_available() else "cpu") model = model.to(device) optimizer = torch.optim.SGD( model.fc.parameters(), # ← HANYA fc layer! lr=0.001, momentum=0.9 ) loss_fn = nn.CrossEntropyLoss()

# =========================== # Fine-Tuning: Unfreeze layer terakhir # =========================== # Load pre-trained model model = models.resnet18(weights='IMAGENET1K_V1') # Freeze SEMUA dulu for param in model.parameters(): param.requires_grad = False # Unfreeze layer4 (layer terakhir sebelum FC) for param in model.layer4.parameters(): param.requires_grad = True # 🔓 Unfreeze! # Ganti FC head num_features = model.fc.in_features model.fc = nn.Linear(num_features, 10) # =========================== # Discriminative Learning Rates! # Layer dekat input → lr kecil (fitur dasar, jangan ubah banyak) # Layer dekat output → lr besar (perlu adaptasi ke task baru) # =========================== optimizer = torch.optim.Adam([ {'params': model.layer4.parameters(), 'lr': 1e-4}, # LR kecil {'params': model.fc.parameters(), 'lr': 1e-3}, # LR normal ]) # Learning rate scheduler: turunkan LR setiap 7 epoch scheduler = torch.optim.lr_scheduler.StepLR( optimizer, step_size=7, gamma=0.1 ) # Training loop (sama seperti biasa) for epoch in range(15): model.train() for images, labels in train_loader: images, labels = images.to(device), labels.to(device) pred = model(images) loss = loss_fn(pred, labels) optimizer.zero_grad() loss.backward() optimizer.step() scheduler.step() # Update LR # Feature Extraction: ~91-94% akurasi # Fine-Tuning: ~96-98% akurasi ← Jauh lebih baik!

# pip install pytorch-grad-cam from pytorch_grad_cam import GradCAM from pytorch_grad_cam.utils.image import show_cam_on_image # Target layer: layer terakhir sebelum FC cam = GradCAM( model=model, target_layers=[model.layer4[-1]] # Conv layer terakhir ) # Generate heatmap input_tensor = transform(image).unsqueeze(0).to(device) grayscale_cam = cam(input_tensor=input_tensor) # Overlay heatmap di atas gambar asli visualization = show_cam_on_image( rgb_image, # Gambar asli (numpy) grayscale_cam[0], # Heatmap use_rgb=True ) # Tampilkan import matplotlib.pyplot as plt plt.imshow(visualization) plt.title("Grad-CAM: Apa yang Dilihat Model?") plt.axis('off') plt.show()