Task Mining Challenge

Run the application locally

Pre-requirements:

• Python 3.11
• Task installed on your machine. For installation details, see here
• Docker installed on your machine if you want to run the application locally with Docker
• To use the Weights & Biases Board, you need to export your WANDB_API_KEY locally. Run export WANDB_API_KEY="your api key". If you have no Weights & Biases account and do not want to create one, I can provide you a key that is ready for sharing on the interview day.
• For the following run the gesture recognition locally, the trigger training locally and trigger pytest locally, you shall following the instruction under the home directory of this project, i.e./your_directory/task_mining_challenge.

Run the Gesture Recognition application locally

1. Create your local virtual environment by running:

task create-venv
task install-dependencies

2. Activate your virtual environment by running source ./venv/bin/activate
3. Run the application locally:

export WANDB_API_KEY="your api key"
task run-local

Then open your browser and go to http://localhost:5000/docs. You will be directed to the SwaggerUI, where you can trigger the API endpoints. The User name and password to access the swaggerUI is defined in ./app/main.py, which are only the dummy user name and password, just for the sake of showing the authorization functionaity. Note: Saving authorization token in the script is not proper approach in practice. As here is the only the dummy pass and user name, and only for showcasing the functionality, so they are included in the source code. In practice, consider use Key Vault or Database for saving this information.

Run the Gesture Recognition application with Docker Container

To run this application with a Docker container locally, simply run the following lines:

# build your docker container locally 
export WANDB_API_KEY="your api key" 
task docker-build
# run the built docker container
task docker-run

Then you can open the localhost by copying http://localhost:5000/docs to your browser. You can then try out the different API calls in the SwaggerUI.

Test the API calls

• To test the /insert_data endpoint, you can find a sample input in test/insert_data.json.
• To test the /inference endpoint, you can find a sample input for the input_data part in test/inference_input.json, you still need to provide the model version to use for the inference. It is suggested to use model version 1, which was trained with 800 epochs and the best model is saved during the training.
• To test /training endpoint, by default in the ./trainning/config.yaml the number of epochs is defined as 100. If you just want to test this oout and make it run faster, you can set this parameter to for example: 10. The trained model and the used MIC/Standard Scaler can be found in ./training/models/ and ./training/data/preprocess_xx_data/.
• To test /delete_model endpoint, you will find that the model and it's MIC/Standard Scaler will be deleted.

Run Training locally

For developers and data scientists who want to run the training locally with fine-tuning of hyperparameters, changing optimizers, applying other technologies, or changing model architecture, you can use the following instructions.

Trigger training locally

• Activate the virtual environment by running the following command:

  source .venv/bin/activate

  python -m training.train --cfg ./training/config.yaml --model_version 2

Trigger pytest locally

1. Activate the virtual environment source ./venv/bin/activate.
2. Run task pytest-run.

Todos for making this application production-ready:

1. Cloud deployment: Move the solution to a cloud platform (e.g., Azure) for better management and scalability.

   • Use Azure Blob Storage for model and data storage
   • Utilize Azure Container Registry for container image management
   • Deploy to Azure Container/Kubernetes Service for improved scaling

2. Implement data versioning to track changes in datasets over time.

3. Adapt model training tracking:

   • Currently using Weights & Biases
   • Consider enterprise-specific limitations or preferred toolsets
   • Integrate with the company's existing infrastructure

4. Code refactoring:

   • Add type hints for function inputs and outputs
   • Improve documentation for all functions
   • Refactor some scripts to make them more well structured. For exampl: ./training/train.py

5. Enhance testing:

   • Add comprehensive unit tests
   • Expand integration test coverage

6. Separate model training and inference:

   • Split into distinct services or platforms
   • Optimize resource allocation for training
   • Ensure consistent inference performance
   • Improve model tracking and versioning

7. Implement CI/CD pipeline:

   • Automate orchestration
   • Include automated testing in the pipeline
   • Streamline cloud deployment
   • Facilitate smooth model releases

8. Security enhancements:

   • Implement proper authentication and authorization
   • Secure API endpoints
   • Manage secrets and credentials securely

9. Monitoring and logging(Toolset: Azure Log Analytics/Datadog):

   • Set up comprehensive logging
   • Implement performance monitoring
   • Create alerts for critical issues

10. Scalability and load balancing:

    • Design for horizontal scaling
    • Implement load balancing for high availability

11. Error handling and resilience:

    • Improve error handling throughout the application
    • Implement retry mechanisms for transient failures