A simple utility to upload a COCO dataset format to custom vision and vice versa. This can be used to backup your custom vision object detection projects into a storage account and restore it later or use AzureML to create a more custom CV model.

Currently the scripts work with Object Detection but can be easily updated to work with Classification.

Install from pip:

pip install coco2customvision

To export a custom vision project to an Azure storage account use the following:

coco2customvision export -sk "<storage_account_key>" -sn <storage_account_name> -sc <storage_account_container_name> -cvk <custom_vision_key> -cve <custom_vision_endpoint> -cvp <custom_vision_project_name> coco_dataset_filename.json

To import a coco dataset that is located in an Azure storage account container into a custom vision project (the project may not exist yet):

coco2customvision import -sk "<storage_account_key>" -sn <storage_account_name> -sc <storage_account_container_name> -cvk <custom_vision_key> -cve <custom_vision_endpoint> -cvp <custom_vision_project_name> coco_dataset_filename.json

You can get the parameters from:

• Custom vision: custom_vision_key, custom_vision_endpoint, custom_vision_project_name
• Azure portal: storage_account_key, storage_account_name, storage_account_container_name

If you want to contribute to this code base, clone the repo and follow these instructions.

Install module in editable/develop mode (-e) and include the development dependencies (the [dev] argument you see) using the following:

pip install -e .[dev]

For tests to complete, you need to configure some secrets. These secrets are retrieved from environment variables. To avoid adding these environment variables in your system, you need to create a pytest.ini file based on the pytest.ini.template template and fill in all needed values. Use the following links to retrieve the corresponding settings:

• Custom vision
• Azure storage account.

The pytest.ini file is in .gitignore to avoid pushing credentials accidentally.

To run all tests:

python -m pytest . -c pytest.ini

or use the VSCode Test Explorer to even debug your code.

If you installed the module in develop mode you can use it directly as seen in the instructions above. You can also use the module reference, as seen bellow.

To export a custom vision project to an Azure storage account use the following:

python -m src.coco2customvision export -sk "<storage_account_key>" -sn <storage_account_name> -sc <storage_account_container_name> -cvk <custom_vision_key> -cve <custom_vision_endpoint> -cvp <custom_vision_project_name> coco_dataset_filename.json

To import a coco dataset that is located in an Azure storage account container into a custom vision project (the project may not exist yet):

python -m src.coco2customvision import -sk "<storage_account_key>" -sn <storage_account_name> -sc <storage_account_container_name> -cvk <custom_vision_key> -cve <custom_vision_endpoint> -cvp <custom_vision_project_name> coco_dataset_filename.json

Before making any commit you can invoke the pre-commit.bat file which does the following:

• Format the code using black:

  python -m black . 

• Ensure that there is no flake8 error:

  python -m flake8 .

• Ensure all test pass:

  python -m pytest . -c pytest.ini

• Ensure setup.cfg file is consistently formatted:

  setup-cfg-fmt setup.cfg

To create a release you need to create an annotated tag:

git tag -a v0.1.0 -m "First version of the tool"

You can view existing tags and their comments (-n) using:

git tag -n

Run a build to create the corresponding version artifacts under the dist folder:

python -m build

Then push them to testpypi to verify:

pip install --upgrade twine
twine upload --repository testpypi dist/*

Verify results in the test Pypi instance. You can try installing in a new python environment using:

pip install --index-url https://test.pypi.org/simple/ --extra-index-url https://pypi.org/simple coco2customvision

If you decide to make some more changes, you can delete a tag using the following:

git tag -d v0.1.0

Note: You will not be able to push the same version to the test Pypi instance. As a work around you can increase the prerelease 4th digit e.g. 0.1.0.1.

When you are ready push changes to remote and let github actions publish the package to the production Pypi. Just push the tag to GitHub and the CD action will create the release:

git push origin --tags

Note that currently the CD process is kicked when you push the tag and it doesn't do the CI part. So make sure your code is passing the CI part before tagging and pushing the tag to GitHub.

Here is a list of related projects and references to this effort:

• Custom vision blob connector python tool to upload images to custom vision from blob storage.
• CustomVision.COCO C# tool to train models using a COCO definition file.
• VoTT2COCO converts VoTT json files to COCO format.
• Azure cognitive services python SDK
• Custom vision SDK
• Description of the COCO format
• Coco2Yolo
• Unofficial tool to download a CVS project in COCO format: https://github.com/shonohs/cvsutils, https://github.com/shonohs/simpledataset

  cvs_download_project <project_id> downloaded/
  dataset_convert_to downloaded/images.txt coco coco_output/

• Custom Vision Autotrainer: Found it after this project was completed. Haven't played with it.

List of Python related references:

• Writing python in VSCode
• Fixtures in pytest
• Click arguments parser
• Packaging projects
• Dependency management
• Docs authoring pack, highly recommended collection of VSCode plugins.