Capstone: End-to-End ML Pipeline

from torchvision import transforms, datasets from torch.utils.data import DataLoader, WeightedRandomSampler # Transforms: augmentasi untuk training, clean untuk val/test train_transform = transforms.Compose([ transforms.RandomResizedCrop(224), transforms.RandomHorizontalFlip(), transforms.RandomVerticalFlip(), transforms.RandomRotation(20), transforms.ColorJitter(brightness=0.2, contrast=0.2), transforms.ToTensor(), transforms.Normalize([.485,.456,.406], [.229,.224,.225]) ]) val_transform = transforms.Compose([ transforms.Resize(256), transforms.CenterCrop(224), transforms.ToTensor(), transforms.Normalize([.485,.456,.406], [.229,.224,.225]) ]) # Weighted sampler untuk handle class imbalance class_weights = 1.0 / torch.tensor(class_counts, dtype=torch.float) sample_weights = class_weights[labels] sampler = WeightedRandomSampler(sample_weights, len(sample_weights)) train_loader = DataLoader(train_set, batch_size=32, sampler=sampler) val_loader = DataLoader(val_set, batch_size=64, shuffle=False)

from torchvision import models model = models.resnet50(weights='IMAGENET1K_V2') # Freeze early layers, unfreeze layer3 + layer4 for name, param in model.named_parameters(): if "layer3" not in name and "layer4" not in name: param.requires_grad = False # Custom classification head model.fc = nn.Sequential( nn.Dropout(0.3), nn.Linear(2048, 512), nn.ReLU(), nn.Dropout(0.2), nn.Linear(512, 7) # 7 skin lesion classes ) # Compile for speed! model = torch.compile(model)

import wandb from torch.cuda.amp import autocast, GradScaler wandb.init(project="skin-cancer-capstone") scaler = GradScaler() optimizer = torch.optim.AdamW([ {'params': model.layer3.parameters(), 'lr': 1e-5}, {'params': model.layer4.parameters(), 'lr': 5e-5}, {'params': model.fc.parameters(), 'lr': 1e-3}, ], weight_decay=0.01) scheduler = torch.optim.lr_scheduler.CosineAnnealingLR( optimizer, T_max=30, eta_min=1e-7 ) best_val_acc = 0 patience, counter = 5, 0 for epoch in range(30): model.train() for imgs, labels in train_loader: imgs, labels = imgs.to(device), labels.to(device) with autocast(device_type="cuda"): loss = loss_fn(model(imgs), labels) scaler.scale(loss).backward() scaler.step(optimizer); scaler.update() optimizer.zero_grad() val_acc = evaluate(model, val_loader) scheduler.step() wandb.log({"val_acc": val_acc, "lr": scheduler.get_last_lr()[0]}) # Early stopping + best model save if val_acc > best_val_acc: best_val_acc = val_acc torch.save(model.state_dict(), "best_model.pt") counter = 0 else: counter += 1 if counter >= patience: break # Hasil: 93.7% accuracy, 0.91 F1-score, 0.89 AUC-ROC

from sklearn.metrics import classification_report, confusion_matrix from pytorch_grad_cam import GradCAM # Classification metrics print(classification_report(y_true, y_pred, target_names=class_names)) # precision recall f1-score # melanoma 0.89 0.87 0.88 # benign_kerat 0.94 0.96 0.95 # dermatofibroma 0.91 0.88 0.89 # ... # accuracy 0.937 # Grad-CAM: model fokus di mana? cam = GradCAM(model=model, target_layers=[model.layer4[-1]]) # → Heatmap menunjukkan model fokus di area lesi kulit ✅ # → Bukan background atau rambut (good sign!)

# 1. Export ke ONNX torch.onnx.export(model, dummy, "skin_classifier.onnx", dynamo=True) # 2. FastAPI endpoint @app.post("/classify") async def classify_lesion(file: UploadFile): img = preprocess(await file.read()) result = session.run(None, {"input": img}) return {"diagnosis": CLASSES[result.argmax()], "confidence": float(result.max())} # 3. Docker: docker build -t skin-api . && docker run -p 8000:8000 skin-api # curl -F "file=@lesion.jpg" http://localhost:8000/classify # → {"diagnosis": "melanoma", "confidence": 0.923}

skin-cancer-classifier/ ├── data/ │ ├── raw/ # HAM10000 original images │ └── processed/ # Split: train/val/test ├── src/ │ ├── data_pipeline.py # Transforms, DataLoader (Part 2) │ ├── model.py # ResNet-50 + custom head (Part 3-4) │ ├── train.py # AMP, scheduler, W&B (Part 9) │ ├── evaluate.py # Metrics, Grad-CAM (Part 4) │ └── api.py # FastAPI server (Part 6) ├── models/ │ ├── best_model.pt # Best checkpoint │ └── model.onnx # ONNX export (Part 6) ├── notebooks/ │ └── exploration.ipynb # Data exploration ├── Dockerfile # Container config (Part 6) ├── requirements.txt ├── README.md └── .github/workflows/ # CI/CD pipeline