This is a template for multi-modal machine learning in healthcare using the Kedro framework. You can combine reports, tabular data and images using various fusion methods (Early & Late fusion. Few other fusion methods and graph data are WIP). This project works with Kubeflow and Vertex AI.

• If you are not familiar with the Kedro platform, please read the overview 👇
• This is a template repository. Generate a new repository with the same directory structure by clicking the Use this template button ☝️ and use it as a Kedro project.
• Install dependenties pip install -r src/requirements.lock
• Refer default pipeline for usage examples.
• Refer sample data for data format. Prefix model datasets with appropriate model type from image_ , text_ , tabular_ , and bert_. (text_ is for CNN text models)
• Refer catalogue for inputs and outputs
• See parameters that can be tweaked.

The required pipelines are in requirements.txt. More details on the components are in their respective repositories 👇 (PR welcome. Read CONTRIBUTING.md in the repositories)

• kedro-tf-image
• kedro-tf-text
• kedro-tf-utils
• kedro-dicom (optional) for processing DICOM images
• kedro-graph (optional) for creating DGL graph from multimodal data.
• fhiry (optional) for flattening FHIR resources.

• Use any number/combination of data types.
• Export trained fusion model for TF serving with an additional signatureDef for receiving image as b64 string. Use SERVING: path/to/save/model in parameters. Also see serving.py (example serving REST client) and serving.sh (start TF serving docker container with the model).
• Experimental support for BERT models with some caveats.

• Please note that you can build any multi-modal architecture!

If you find this project useful, give us a star. It helps others discover the project.

• Downloaded BERT models will not copy vocab.txt in assets folder to the newly created fusion model. This has to be manually copied.
• The class_num in TfModelWeights must be equal to to NCLASSES during training. Otherwise it throws an error: Tensorflow estimator ValueError: logits and labels must have the same shape ((?, 1) vs (?,))
• import tensorflow_text as text is required in pipeline_registry to avoid this error: Op type not registered 'CaseFoldUTF8' in binary running on xxxx
• Inconsistent NCLASSES leads to ValueError: logits and labels must have the same shape, received ((None, 3) vs (None, 1)).

• Bell Eapen |

This is your new Kedro project, which was generated using Kedro 0.18.4.

Take a look at the Kedro documentation to get started.

In order to get the best out of the template:

• Don't remove any lines from the .gitignore file we provide
• Make sure your results can be reproduced by following a data engineering convention
• Don't commit data to your repository
• Don't commit any credentials or your local configuration to your repository. Keep all your credentials and local configuration in conf/local/

Declare any dependencies in src/requirements.txt for pip installation and src/environment.yml for conda installation.

To install them, run:

pip install -r src/requirements.txt

You can run your Kedro project with:

kedro run

Have a look at the file src/tests/test_run.py for instructions on how to write your tests. You can run your tests as follows:

kedro test

To configure the coverage threshold, go to the .coveragerc file.

To generate or update the dependency requirements for your project:

kedro build-reqs

This will pip-compile the contents of src/requirements.txt into a new file src/requirements.lock. You can see the output of the resolution by opening src/requirements.lock.

After this, if you'd like to update your project requirements, please update src/requirements.txt and re-run kedro build-reqs.

Further information about project dependencies

Note: Using kedro jupyter or kedro ipython to run your notebook provides these variables in scope: context, catalog, and startup_error.

Jupyter, JupyterLab, and IPython are already included in the project requirements by default, so once you have run pip install -r src/requirements.txt you will not need to take any extra steps before you use them.

To use Jupyter notebooks in your Kedro project, you need to install Jupyter:

pip install jupyter

After installing Jupyter, you can start a local notebook server:

kedro jupyter notebook

To use JupyterLab, you need to install it:

pip install jupyterlab

You can also start JupyterLab:

kedro jupyter lab

And if you want to run an IPython session:

kedro ipython

You can move notebook code over into a Kedro project structure using a mixture of cell tagging and Kedro CLI commands.

By adding the node tag to a cell and running the command below, the cell's source code will be copied over to a Python file within src/<package_name>/nodes/:

kedro jupyter convert <filepath_to_my_notebook>

Note: The name of the Python file matches the name of the original notebook.

Alternatively, you may want to transform all your notebooks in one go. Run the following command to convert all notebook files found in the project root directory and under any of its sub-folders:

kedro jupyter convert --all

To automatically strip out all output cell contents before committing to git, you can run kedro activate-nbstripout. This will add a hook in .git/config which will run nbstripout before anything is committed to git.

Note: Your output cells will be retained locally.

Further information about building project documentation and packaging your project