Brain Tumor 🧠 Classification with Docker Compose 🐳 and PyTorch Lightning ⚡

This project demonstrates how to set up training, evaluation, and inference for Brain Tumor Classification using Docker and PyTorch Lightning. We employ Docker containers for environment consistency, and PyTorch Lightning for structured, modular training, evaluation, and inference processes. The dataset used in this project is the Brain Tumor Classification (MRI) dataset, managed through Docker Compose.

Table of Contents

• Requirements
• Dataset
• PyTorch Lightning Module
  • Why PyTorch Lightning?
• Docker Setup
  • DevContainer
  • Docker Compose Services
• Training and Evaluation
• Inference
• Running Docker Containers
• Results
• References

Requirements 📦

This project requires the following packages to be installed:

• PyTorch
• torchvision
• PyTorch Lightning
• NumPy
• scikit-learn
• TensorBoard
• Matplotlib

Dataset 📂

The dataset used in this project is the Brain Tumor Classification (MRI). It consists of MRI scans categorized into four classes: 👇

Glioma Tumor	Meningioma Tumor
Normal	Pituitary Tumor

Each class is stored in separate directories, allowing for easy management and access during the training and evaluation process.

PyTorch Lightning Module ⚡

PyTorch Lightning simplifies the training process by modularizing the code and separating concerns. It helps you focus more on the logic of your models instead of boilerplate code.

Why PyTorch Lightning?

✅ Simplified Code Structure: It encourages clean and organized code, making it easier to maintain and scale.

✅ Flexibility: It allows for an easy switch between different training strategies (such as multi-GPU training, TPU support, etc.) with minimal changes.

✅ Built-in Features: It includes built-in logging, checkpointing, and early stopping, reducing manual implementations.

Docker Setup 🐳

DevContainer

The .devcontainer setup allows you to develop in a pre-configured environment using VS Code:

{
  "name": "Brain Tumor Classification",
  "dockerFile": "Dockerfile",
  "settings": {
    "terminal.integrated.shell.linux": "/bin/bash"
  },
  "extensions": [
    "ms-python.python",
    "ms-azuretools.vscode-docker"
  ],
  "postCreateCommand": "pip install -r requirements.txt"
}

Docker Compose Services

The Docker Compose file defines three services: train, evaluate, and infer:

version: '3.8'

services:
  train:
    # Train service
    build:
      context: .
      dockerfile: Dockerfile.train
    shm_size: "2gb"
    volumes:
      - host:/opt/mount
      - ./model:/opt/mount/model
      - ./data:/opt/mount/data

  evaluate:
    # Evaluate service
    build:
      context: .
      dockerfile: Dockerfile.eval
    volumes:
      - host:/opt/mount
      - ./model:/opt/mount/model
      - ./data:/opt/mount/data

  infer:
    # Inference service
    build:
      context: .
      dockerfile: Dockerfile.infer
    volumes:
      - host:/opt/mount
      - ./model:/opt/mount/model
      - ./data:/opt/mount/data
      - ./predictions:/opt/mount/results

volumes:
  host:

Training and Evaluation

To set up and execute the training, evaluation, and inference processes, follow these steps:

1️⃣ Build Docker Images:

• Build the Docker images for all services using:

  docker compose build

2️⃣ Train the Model:

• To train the model, run:

  docker compose run train

• The BrainTumorClassifier will train the model and save the checkpoints in the shared volume.

3️⃣ Evaluate the Model:

• To evaluate the model, execute:

  docker compose run evaluate

• This will load the saved checkpoint and print the validation metrics using eval.py.

Inference 🔍

To run inference on sample images:

docker compose run infer

• The infer.py script will predict the labels for 10 random images and save the output to the predictions directory.

Each service uses volume mounts to ensure that data, checkpoints, and results are accessible across different containers, maintaining consistency.

Results 📊

Check the predictions in the predictions folder. Below are some sample results:

Actual: Glioma Tumor Predicted: Glioma Tumor (Confidence: 1.00)	Actual: Glioma Tumor Predicted: Glioma Tumor (Confidence: 1.00)
Actual: Meningioma Tumor Predicted: Meningioma Tumor (Confidence: 1.00)	Actual: Meningioma Tumor Predicted: Meningioma Tumor (Confidence: 0.81)
Actual: Normal Predicted: Normal (Confidence: 0.98)	Actual: Pituitary Tumor Predicted: Pituitary Tumor (Confidence: 0.99)
Actual: Pituitary Tumor Predicted: Pituitary Tumor (Confidence: 0.51)	Actual: Pituitary Tumor Predicted: Meningioma Tumor (Confidence: 0.43)

References 🔗

• Brain Tumor Classification (MRI) Dataset on Kaggle
• PyTorch
• PyTorch Lightning
• Docker Documentation