WebCards Documentation

Welcome to the documentation of the WebCards tool! Here you can find useful information about the app, its logic, and its use. WebCards is a tool for creating and sharing model cards and datasheets, which are documents that provide context and transparency for machine learning tasks. The tool is under development and heavy testing. If you find any bugs or have any suggestions, please create an issue in GitHub page.

Overview

Built with the Django framework, this web application leverages Google’s Model Card Toolkit as a basis for creating model cards. A customized library is created which can be found here.

The application allows users to input data and subsequently generates a model card as an HTML file. Users can upload datasets, model and graph files, which are then processed and incorporated into the model card. The uploaded files are temporarily stored in the media/uploads directory and are associated with a session key that’s stored in the browser’s cookies. Users have the option to delete these files, otherwise, they will be automatically removed after one hour.

We use SQLite3 as our database system due to its simplicity and speed, which are adequate for the current scale and complexity of our tool.

A CI/CD pipeline has been implemented using GitHub Actions for testing, building, and pushing updates to DockerHub. This ensures that data doesn’t accumulate in the system, as files are either overwritten or deleted after usage.

Getting Started

The tool is currently in the early stages of production. Once it’s ready for public use, the link will be provided here.

To deploy the tool locally, follow these steps:

  1. Clone the Repository:

    git clone https://github.com/mcmi-group/feat_ai.git

  2. Create a New Virtual Environment:

    conda create --name webcards python=3.8

  3. Activate the Virtual Environment:

    conda activate webcards

  4. Navigate to the Web App Directory:

    cd rai_webapp

  5. Install the Required Packages:

    pip install -r requirements.txt

  6. Install the Model Card Toolkit:

    pip install --upgrade utils/model_card_toolkit-2.0.0.dev0-py3-none-any.whl

  7. Run the Server:

    unzip db.sqlite3.zip

python manage.py runserver

After following these steps, you should have the tool set up and ready to run locally.

Technical Overview

Currently, the tool can generate model cards and datasheets. Central to the model card generation is the ModelCard dataclass, which serves as a blueprint for the entire card. This dataclass is structured hierarchically:

ModelCard Dataclass
│
├── Section 1 (Subclass of ModelCard)
│   ├── Field 1.1 (Subclass of Field)
│   │   ├── Subfield 1.1.1 (Subclass of Field 1.1)
│   │   ├── Subfield 1.1.2
│   │   └── ...
│   │
│   ├── Field 1.2
│   └── ...
│
├── Section 2
│   ├── Field 2.1 (Subclass of Field)
│   │   ├── Subfield 2.1.1 (Subclass of Field 2.1)
│   │   └── ...
│   │
│   ├── Field 2.2
│   └── ...
│
└── ...

Each section in the ModelCard dataclass contains fields, which can further have subfields, allowing for a detailed and granular representation of the model card’s content.

For datasheets, the process is streamlined. User responses are directly rendered into HTML, bypassing the dataclass representation.

User inputs, collected through the website, are populated into these data structures. The ModelCardGenerator then processes these inputs, culminating in the creation of a model card ready for export.

Through the documentation especially for the developer guide, it is recommended to check the official django documentation for more information about the framework. The documentation can be found here.

Developer Guide

User Guide