System Architecture¶

Overview of the MLOps Skin Lesion Classification Pipeline architecture.

High-Level Architecture¶

System Architecture

Model Architecture¶

We use EfficientNet with transfer learning for skin lesion classification:

Backbone: EfficientNet-B0/B3 pretrained on ImageNet
Classification Head: Custom fully-connected layers for binary classification
Training: Fine-tuning with PyTorch Lightning

Training Pipeline¶

Training Pipeline

Data Pipeline¶

graph LR A[HAM10000 Dataset] --> B[DVC] B --> C[Google Cloud Storage] C --> D[Local Cache] D --> E[DataLoader] E --> F[Model Training]

Inference Pipeline¶

graph LR A[Client Request] --> B[FastAPI] B --> C[Image Preprocessing] C --> D[ONNX Runtime] D --> E[Postprocessing] E --> F[JSON Response]

Deployment Architecture¶

graph TD A[GitHub] --> B[GitHub Actions CI/CD] B --> C[Docker Build] C --> D[Container Registry] D --> E[Google Cloud Run] E --> F[Load Balancer] F --> G[Users]

Technology Components¶

Component	Technology	Purpose
ML Framework	PyTorch Lightning	Training orchestration
Configuration	Hydra	Experiment configs
Data Versioning	DVC	Dataset management
Experiment Tracking	Weights & Biases	Metrics logging
API Server	FastAPI	Inference endpoint
Model Runtime	ONNX Runtime	Fast inference
Containerization	Docker	Reproducible deployment
Cloud Platform	Google Cloud Run	Serverless hosting
CI/CD	GitHub Actions	Automated testing

Directory Structure¶

dtu-vibe-ops-02476/
├── src/mlops_project/     # Source code
│   ├── model.py           # EfficientNet model
│   ├── data.py            # Dataset handling
│   ├── train.py           # Training script
│   ├── evaluate.py        # Evaluation
│   └── api.py             # FastAPI server
├── configs/               # Hydra configurations
├── data/                  # Dataset (DVC managed)
├── models/                # Trained models
├── tests/                 # Test suite
├── dockerfiles/           # Container definitions
└── docs/                  # Documentation