Predicting probability of defaul

Workflow

• Model training
• Generate predictions
• Deploy the model to AWS using SageMaker

Project Structure

|-- data
    |-- submission.csv
|-- sagemaker
     |-- data
     |-- models
     |-- AWS - SageMaker.ipynb
     |-- featurizer_local.py
     |-- featurizer_remote.py
|-- src
    |-- models
        |-- baseline_model.py: Class for training the binary classification model
        |-- transformers.py: ColumnSelector transformer for preprocessing
|-- api_request_exmpl.json
|-- pickled_model.pickle
|-- Model.ipunb: model evaluation notebook

Pre-requisites

• python3.8
• aws-cli

Submissions

1. A verbose explanation of the model training and validation could be found here.
2. CSV file with resulting predictions could be found here.

Steps

Model training (/Model.ipynb)

1. Read and explore dataset (excluded from the current repo due to privacy reasons)
2. Train baseline defualt prediction model
3. Evaluate and сompare different models (Logistic Regression and Tree-based ensemble models) with a set of metrics
4. Estimate feature inportance
5. Generate predictions (/data/submission.csv)
6. Dump the best estimator (/pickled_model.pickle)

Model deployment (sagemaker/AWS - SageMaker.ipynb)

1. Define Sagemaker session and role
2. Preprocessing data and train the model
3. Create SageMaker Scikit Estimator
4. Batch transform training data
5. Fit a Tree-based Model with the preprocessed data
6. Serial Inference Pipeline with Scikit preprocessor and classifier
7. Deploy model
8. Make a request to the pipeline endpoint

API Gateway endpoint exposing

Besides AWS SageMaker I tested some other AWS deployment options including AWS Serverless Application Model (SAM). It allows deploing ML models with Serverless API (AWS Lambda). I managed to expose API endpoint with helloworld application behind in order to try AWS Lambda.

Used technologies

• ECR: Container & Registy
• AWS Lambda: Serving API
• SAM: Serverless Framework

Query the endpoint

GET request:

https://x3jp27x3t3.execute-api.eu-central-1.amazonaws.com/test/hello

Expected response:

{
    "message": "hello world"
}