swint-trans-unet/trans-unet

Welcome to Streamlit!

Edit /src/streamlit_app.py to customize this app to your heart's desire. :heart:

If you have any questions, checkout our documentation and community forums.

Swin-Unet ISIC Segmentation

This is a Streamlit application for testing the Swin-Unet model on ISIC dataset for skin lesion segmentation.

Setup

1. 1.Install dependencies:

pip install -r requirements.txt

1. 1.Run the application:

streamlit run src/streamlit_app.py

Usage

1. 1.Select the image size and batch size from the sidebar
2. 2.Choose the ISIC dataset version (2018/2019/2020)
3. 3.Click "Start Testing" to run the model on the test set
4. 4.View the results including:
5. 5.Dice Score
6. 6.Accuracy
7. 7.IoU
8. 8.Sample predictions with visualizations

Model

The application uses a pre-trained Swin-Unet model from Hugging Face (K0ZUME/best_model.pth).

Dataset

The application uses the ISIC dataset from Hugging Face, which can be selected from:

• —ISIC 2018
• —ISIC 2019
• —ISIC 2020

TransUNet with ISIC Dataset Training

This repository contains a Streamlit-based implementation of TransUNet for training on the ISIC dataset. The implementation is based on the Swin-Unet architecture and includes a user-friendly interface for training and monitoring the model's progress.

Features

• —Streamlit-based training interface
• —Real-time training progress monitoring
• —Support for ISIC dataset
• —Configurable training parameters
• —Model checkpointing
• —TensorBoard integration for visualization

Installation

1. 1.Clone the repository:

git clone https://huggingface.co/spaces/your-username/trans-unet
cd trans-unet

1. 1.Install the required dependencies:

pip install -r requirements.txt

Data Preparation

1. 1.Download the ISIC dataset and prepare it in the following format:
2. 2.Each image and its corresponding mask should be saved as a .npz file
3. 3.The .npz file should contain two arrays: 'image' and 'label'
4. 4.Create train.txt, val.txt, and test.txt files listing the filenames (without .npz extension)

1. 1.Place the data in the following structure:

data/
  ISIC/
    train_npz/
      image1.npz
      image2.npz
      ...
lists/
  lists_ISIC/
    train.txt
    val.txt
    test.txt

Usage

1. 1.Start the Streamlit app:

streamlit run src/streamlit_app.py

1. 1.Configure the training parameters in the sidebar:
2. 2.Image Size: Size of input images (default: 224)
3. 3.Batch Size: Number of images per batch (default: 24)
4. 4.Max Epochs: Maximum number of training epochs (default: 150)
5. 5.Learning Rate: Initial learning rate (default: 0.01)
6. 6.Data Root Path: Path to the ISIC dataset
7. 7.List Directory: Path to the directory containing train.txt, val.txt, and test.txt
8. 8.Output Directory: Directory to save model checkpoints and logs
9. 9.Config Path: Path to the model configuration file

1. 1.Click "Start Training" to begin the training process

1. 1.Monitor the training progress:
2. 2.Progress bar shows overall training progress
3. 3.Training and validation losses are displayed in real-time
4. 4.Best model checkpoints are saved automatically

Model Architecture

The model is based on the Swin Transformer architecture with the following key components:

• —Patch Embedding Layer
• —Swin Transformer Blocks
• —Skip Connections
• —Final Classification Head

Configuration

The model configuration can be modified in configs/swin_tiny_patch4_window7_224_lite.yaml. Key parameters include:

• —Model architecture (embedding dimension, number of layers, etc.)
• —Training parameters (batch size, learning rate, etc.)
• —Data augmentation settings
• —Loss function configuration

License

This project is licensed under the MIT License - see the LICENSE file for details.

Acknowledgments

• —The implementation is based on the Swin-Unet architecture
• —ISIC dataset for providing the medical image segmentation data