Deployable ML model with Flask and Docker, based on housing data.
A system with Docker installed.
Clone the repo to your machine of choice, and using your command line emulator, enter the top level directory and run:
docker-compose up --buildThis should build the Docker container and start the REST server at http://localhost:5000.
If you prefer to change this port, simply edit the compose.yaml file and specify the port of your
choice, and edit the Dockerfile to expose that port.
There is a homepage that is accessible via web browser at http://localhost:5000/.
The homepage includes usage examples for making predictions via the REST endpoints, /predict and
/predict_lgbm.
Additionally, the test scripts in ./sound-realty/tests/test_predict.sh and
./sound-realty/tests/test_predict.sh contain more context that can be used to make predictions via
the API.
Input data to the API should be provided in a JSON format, where keys correspond to a specific set of features, and values correspond to one sample from the housing dataset. The expected features are:
['bedrooms', 'bathrooms', 'sqft_living', 'sqft_lot', 'floors', 'sqft_above', 'sqft_basement', 'zipcode']The feature zipcode is used to add various location-based features on the backend but is not used directly in the model.
The /predict endpoint implements a KNN model which follows a Scikit-Learn RobustScaler in a pipeline.
The API can be called from a terminal using curl as follows:
curl -d @./data/test_row.json -H "Content-Type: application/json" http://localhost:5000/predictwhere test_row.json can be found in the data/ directory.
The /predict_lgbm endpoint uses a LightGBM
model and a set of custom, derived features to predict housing prices.
The LightGBM model improves upon the KNN method above in terms of mean absolute precision error
(MAPE) performance, with respect to the KNN technique.
It can be called as above, by simply changing the endpoint.
Please refer to the Jupyter Notebook in model_training.ipynb for more information on how the LightGBM model is trained. It should be noted that this notebook was hosted and trained in a different environment than the Docker container for the overall repo. So if you wish to run that code, you will need to install additional dependencies, or stand up a local virtual environment.
Refer to Notes.md for more information.