Train a machine learning (ML) model based on the data from Titanic disaster and apply it to calculate the survival rate for passenger of the Titanic 2. Details of this challenge and data set for ML model trainig can be found here.
Train data preparation and ML model training was done using Jupyter notebook titanic2_train.ipynb. PassengerId, Name, Ticket and Cabin columns from the original data set were not used for model training. Several columns (Pclass, Sex, Embarked) have been quantified or normalized for training. Survived column was used as a data flag. TensorFlow Sequential model consisted of four fully connected layers was used. In the last layer sigmoid activation function was used to obtain result in the range from 0 to 1. Trained model was saved in the ML_model_titanic2 directory. Finally directory with ML model was packed to archive by:
tar -cvzf ML_model.tar.gz ML_model_titanic2
REST API was developed using Flask and saved as app.py file. It consists of two requests:
- GET request at localhost:5000/example which returns json file which can be used as a template for POST request with passenger data
- POST request at localhost:5000/predict which will accept json file with passenger data and return json file with calculated survival rate, to calculate the survival rate prediction function from the calculate.py script will be called
Blueprint for json with the sample input data:
{
"Age": 38,
"Pclass": 2,
"SibSp": 0,
"Parch": 1,
"Fare": 30.00,
"Sex": "male",
"Embarked": "S"
}
Dockerfile was created to build the Docker image. All neccessary files were copied to the work direcory (in case of the ML model archive ADD method was used to unpack it). Library dependencies were defined in the dependencies.txt file. Image was build from inside of the Visual Studio Code with Docker extension and saved as titanic2_restapi. Docker documentation about build process can be found here.
Docker image can be run as a container using the following command:
docker run -p 5000:5000 titanic2_restapi
Test request can be send by running the titanic_request.py script.