Deep Learning Final Project Group 11

System Requirements

• Operating System: Linux Ubuntu 22.04.5
• Python Version: 3.10
• Anaconda: Installed
• PyTorch Version: 2.5.1
• CUDA Version: 12.4
• Other Packages: pandas, numpy, einops, albumentations, imbalanced-learn, optuna, scipy, tqdm, etc.

How to Run

1. Create and Activate Conda Environment

First, create a new Conda environment and activate it:

conda create -n ml_project python=3.11
conda activate ml_project

2. Install Required Packages

Install the required Python packages in the activated Conda environment:

Install Basic Packages

conda install pandas numpy scikit-learn scipy tqdm -y

Install PyTorch

conda install pytorch torchvision torchaudio pytorch-cuda=12.4 -c pytorch -c nvidia

Install Other Required Packages

pip install imbalanced-learn optuna

3. Dataset

import kagglehub

# Download latest version
path = kagglehub.dataset_download("whats2000/breast-cancer-semantic-segmentation-bcss")

print("Path to dataset files:", path)

Dataset Download Link

4. Data Preparation

Place the dataset in the 'archive' folder under the project directory, including the following files: The folder structure should be as follows:

project_directory/
├── model_13.ipynb
├── archive/
│   ├── train_512
│   ├── train_mask_512
│   └── val_512
│   └── val_mask_512
└── bcss_dataset.py

5. Run the Main Program

Execute model_13.ipynb in Jupyter Notebook

The program will automatically execute the following steps:

1. Data Loading and Preprocessing:

   • Load training and validation datasets using BCSSDataset class
   • Perform data augmentation using albumentations, including:
     • Random cropping (320x320)
     • Horizontal and vertical flips
     • 90-degree rotation
     • Brightness and contrast adjustment
     • Hue, saturation, and value adjustment
     • Gaussian noise, Gaussian blur, motion blur
     • Normalization (mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])

2. Model Architecture Design: The program implements 6 different deep learning model architectures:

   • UNet: Classic encoder-decoder architecture
   • nnUNet: Improved UNet with instance normalization and residual connections
   • MultiResUNet: UNet with multi-resolution feature extraction
   • TransUNet: Hybrid architecture combining Transformer and UNet
   • SwinUNet: UNet variant using Swin Transformer
   • DenseUNet: UNet architecture incorporating DenseNet features

3. Training and Optimization Process:

   • Uses AdamW optimizer
   • Implements CosineAnnealingLR learning rate scheduler
   • Implements early stopping mechanism with patience of 8 epochs
   • Uses cross-entropy loss function
   • Monitors multiple evaluation metrics during training:
     • IoU (Intersection over Union)
     • Dice coefficient
     • Accuracy

4. Model Evaluation and Comparison:

   • Records and compares for each model:
     • Training and validation loss
     • IoU score
     • Dice coefficient
     • Accuracy
     • Training time
     • Model parameter count

5. Visualization and Results Analysis:

   • Plots learning curve comparison graphs
   • Generates model performance comparison tables
   • Creates radar charts for multi-dimensional comparison
   • Exports comparison results to CSV file

6. View Results

After training completion, the training process and final model performance visualizations will be displayed in the last cell.

Important Notes

• GPU Support: Ensure that an NVIDIA GPU is installed and CUDA is properly configured to accelerate model training. Use the nvidia-smi command to check GPU status.
• Resource Requirements: Training deep learning models may require substantial memory and computational resources. It's recommended to use a machine with sufficient GPU memory.
• Data Path: If data is located in a different path, modify the data path settings in the main program to ensure correct data loading.
• Dependency Versions: Ensure all package versions match this guide to avoid dependency conflicts. Pay special attention to PyTorch and CUDA version compatibility.