Time-Travelling Visualizer 🚀

A VS Code extension for visualizing and debugging the training process of machine learning models.

✨ Features

• Training Visualization: Visualize the training process through dimensionality reduction.
• Time-Travelling Debugging: Step through training events to understand model behavior over time.
• Influence Analysis: Inspect influence functions to see how training data affects predictions.
• Interactive Tools: Exploring the Representation Space with Multiple Functions.

🛠️ Setup & Launch

1. Setup Environment

   • Backend (Python): Run the setup script below (make sure you have installed Anaconda or Miniconda):

     bash setup.sh

     This will create a Python 3.10 environment named "visualizer" and install the required dependencies.

   • Frontend (Node.js): We assume you have Node.js installed. Navigate to the web and extension directories to install the dependencies:

     # Install web dependencies
     cd web
     npm install
     cd ..
     
     # Install extension dependencies
     cd extension
     npm install

2. Launch

   • Backend Server:
     1. Open a terminal and navigate to the project root directory.
     2. Activate the Conda environment:

        conda activate visualizer

     3. Start the backend server:

        cd tool/server
        python server.py

   • Frontend (Run Extension):
     1. Open the project root directory In VS Code, press Ctrl+Shift+D or click the "Run and Debug" icon in the sidebar.
     2. Select the "Run Extension" configuration from the dropdown menu.
     3. Press F5 or click the "Start Debugging" button.
     4. This will open a new "Extension Development Host" window where the extension is running.

💭 This is only a temporary way to start up the extension; the final version will switch to a regular VSCode extension that can be imported and used directly. For detailed usage instructions, see Doc for time-travelling-visualizer.

📁 Project Structure

.
├── extension/    # 📦 VS Code Extension Source Code
├── tool/         # 🐍 Python Backend & Visualization Methods
├── web/          # 🌐 Frontend Web Application (Loaded In Extension)
├── setup.sh      # 🚀 Setup Script
└── README.md

🚀 Future Improvements

1. Code Refactoring:
   • Decouple various visualization algorithm modules.
   • Elegantly encapsulate the communication between the extension, web page, and backend.
   • Refactor and resolve code smells.
2. Broader Task Compatibility:
   • Extend support to more training tasks, such as generative and contrastive learning tasks.
3. Features in Development:
   • The Token Panel and Influence Function features are currently under development.

---