These are instructions on how to submit a model for listing on Modelplace.AI, the AI model marketplace.

Modelplace currently supports PyTorch, Tensorflow, OpenVINO and ONNXRuntime backends.

How Do I submit a model to Modelplace?

To provide a model we ask you to do the following:

1. Provide required model publishing information about yourself and each model you want to publish
2. Set up environment and try our example model (PyTorch Faster R-CNN)
3. Implement your model in a similar way to the given example model (PyTorch Faster R-CNN)
4. Create a Python wheel package with your model
5. Send your model to us with the above information and package you have created

Model publishing information

Required

• Model description - The whole description of what the model does. Example: See the example model page - this is the text under the "Summary" heading.
• Model description preview short - The text people see on the "card" for each model on the model list page. A short (one sentence) description of your model.
• Model name (Full) - Shown on the individual model page (example model page), e.g. Faster region-based convolutional neural network with ResNet-50 FPN backbone
• Model name (Short) - A much shorter model name shown on the model list page, e.g. Faster R-CNN
• Dataset name - Specify if your model uses public data e.g. MSCOCO or Open images, or a custom dataset
• Preview image - we will run a model on the image and will use its output for visualization. This image will be both on the model list pages and on the model page
• License - License type e.g. Apache 2.0, MIT, Proprietary, etc.
• Number of classes - For classifier models, how many distinct classes does it detect?
• Author - Your name or organization name (e.g. University)
• Metrics - The accuracy of your model

Optional

• Dataset link - The link to a dataset that you use for training and testing of your model
• Homepage link - The link to your project page, source code or website for this model
• System requirements - What system setup (e.g. CPU, GPU, and RAM specifications) we should use to serve your model
• Logo - The logo will be added to model previews on listing pages and individual model pages
  Note: This should be svg (png is possible) image with circle shape and size 128x128 pixels at least
• Avatar image - Shown next to your name or organization name on the model listing and indivudal model pages (by default, shows OpenCV logo)
  Note: This should be svg (png is possible) image with circle shape and size 128x128 pixels at least

Environment Setup

• Install venv
  python3.7 -m pip install virtualenv
• Create an empty virtual environment for python3.7
  python3.7 -m virtualenv venv
• Activate it
  source venv/bin/activate
• Install pytest and wheel
  python3.7 -m pip install pytest wheel
• Install git and git-lfs
  sudo apt install git git-lfs
• Clone the repo
  git clone https://github.com/opencv-ai/modelplace.git

Example Model

For package and style guidelines see the example package (Faster R-CNN).

• Change directory to the package folder (pytorch_fastercnn) cd modelplace/pytorch_fastercnn
• Install the package python3.7 setup.py bdist_wheel && rm -R build/ *.egg-info && pip3 install dist/*.whl
• Run tests python3.7 -m pytest

To see how this example model is represented on the site, view its page: https://modelplace.ai/models/2

Packaging Models for Modelplace

Modelplace uses Python Wheels to simplify model serving. You should create a wheel package with our interfaces. See the example (Faster R-CNN).

To create a package, follow these steps:

• Copy the template folder
• Extend model.py as shown in the example
• Update the setup.py
• Copy your checkpoints to the template/checkpoints folder
• Rename "template" to your package name in all locations, including the folder name, e.g. template -> pytorch_fastercnn
• Install the package locally
  python3.7 setup.py bdist_wheel && rm -R build/ *.egg-info && pip3 install dist/*.whl
• Update tests in test_model.py
• Run tests
  python3.7 -m pytest
• Make sure that the tests are working correctly and all tests pass

Send Your Model To Us

Once you've done all that, to submit your package simply:

• Zip the folder you created in the above steps
• Send us this zip and information about package as described above

That's it! If anything is unclear, please contact us and we'll help you along.