Note

Normally the inference and training pipelines should be separated. I bundled them together into a single repo and created the docker-compose only for the sake of the exercise

Building the container

In order to build the containers please run docker-compose build

Running the container

Please run the following command: docker-compose up This will:

• start a local mlflow server
• run the model-training pipeline that trains the model and exports it to MLFlow
• start a FastAPI based inference container

You should now be able to see a logdir folder. That folder will contain mlflow artifacts and backend as well as a validation-data folder.

That folder contains JSON files with the validation dataset. x_val.json file is structured in a way that makes it easy to test the model with postman/curl.

Calling the API

127.0.0.1:8000/predict

The FastAPI runs on a localhost at port 8000. It has a POST /predict endpoint that takes a JSON body in the following format:

{
    "user_id": "123",
    "features": {
    "age": 52,
    "income": 105824,
    "employment_type": "part_time",
    "marital_status": "married",
    "time_spent_on_platform": 87.6188186,
    "number_of_sessions": 6,
    "fields_filled_percentage": 0.8534444659,
    "previous_year_filing": 0,
    "device_type": "desktop",
    "referral_source": "social_media_ad"
  }
}

Grab any of the data points from the X_val.json file. You can use the index as the user_id

Alternatively you can simply use curl

curl --location '127.0.0.1:8000/predict' \
--header 'Content-Type: application/json' \
--data '{
    "user_id": "123",
    "features": {
    "age": 52,
    "income": 105824,
    "employment_type": "part_time",
    "marital_status": "married",
    "time_spent_on_platform": 87.6188186,
    "number_of_sessions": 6,
    "fields_filled_percentage": 0.8534444659,
    "previous_year_filing": 0,
    "device_type": "desktop",
    "referral_source": "social_media_ad"
  }
}'

Accessing the model info in MLFlow

You should be able to view the training info on http://127.0.0.1:5000/ The newly trained model will be registered as "taxfix_classifier" with a "champion" alias