Censuse data analyes

This project has been written for Udaicty degree in DevOps. It uses Random Forest Classifier to train and predict whether a sample person will earn above or below $50K per year based on cencus data used to train the model.

Built With

Technologies used in the project

• dvc pipeline
• GitHub
• AWS S3
• FastAPI
• Heroku

Getting Started

Conda environment

This project was setup using miniconda. To setup the enviornment run: conda env create -f census_fastAPI.yml

Model documentation

This README defines how to run the project. Model Card tells more about the model used and metrics that comes with the model. Screenshots contains visuals from running the project from command line

Model execution

starter
  EDA.ipynb
  data/
  model/
  metrics/
  params.yaml
  dvc.yaml
  ml/

This project is built with dvc pipeline. Stages are defined in dvc.yaml. To visualise stages use dvc dag:

               +-------------------------+
               | data/clean_data.csv.dvc |
               +-------------------------+
                            *
                            *
                            *
                     +------------+
                     | split_data |***
                   **+------------+   ******
              *****        *                ******
           ***            *                       ******
        ***               *                             ******
+-------+         +--------------+                            ****
| kfold |         | process_data |                               *
+-------+         +--------------+                               *
                   **           **                               *
                 **               **                             *
               **                   **                           *
    +---------------+                 **                       ***
    | train_predict |                  *                   ****
    +---------------+***               *               ****
                        *****          *          *****
                             ****      *      ****
                                 ***   *   ***
                              +---------------+
                              | slice_predict |
                              +---------------+

The original dataset stored under data/census.csv is cleaned with jupyter-notebook EDA.jpynb. All the other stages are executed with dvc repro.

Stages for dvc pipeline

> dvc stage list
split_data     Outputs data/test_data.csv, data/trainval_data.csv
process_data   Outputs data/X.npy, data/y.npy, model/encoder.pkl, model/lb.pkl
kfold          Reports metrics/kfold_scores.json
train_predict  Outputs model/model.pkl; Reports metrics/model_scores.json, metrics/precision.j…
slice_predict  Reports metrics/slice_scores.json

First stage is split_data. All the others depend on it. Stage called kfold can be executed independently as it stores only metrics Stages process_data has to predecess train_predict. Stage slice_predict depends on the model that is stored in stage train_predict.

Artifacts are stored by dvc to s3 remote storage. During CI phase executed by GitHub and CD phase executed by Heroku, the artifacts are pulled from storage.

Currently dvc repro does take all configuration from params.yaml. In next version, there will be a possibility to overwrite arguments from command line.

About `ml/' folder

Script functions.py have functions that are used in pipeline stages. All the other scripts in this folder represent pipeline stages.

Model metrics

All metrics are stored in metrics folder in json format Use dvc exp show to show model performance. For short dvc metrics show will do. More about the metrics you can find in Model Card

Model tests

tests
  conftest.py
  test_fastapi.py
  test_model.py
  sanitycheck.py

Tests are executed with pytest command as part of GitHu workflow. There are 7 tests implemented currently and the config is stored in conftest.py. Script called sanitycheck.py is provided by Udacity and it's used to check whether FastAPI is being properly tested with both POST and GET requests. It can be run manually with python command.

Heroku config files

Aptfile
Procfile
requirements.txt
runtime.txt

Continuouse Delivery is being done on Heroku with FastAPI providing web interface.

FastAPI

main.py
fastAPI_request_scripts

File main.py is defyining our web interface with FastAPI. Python scrips can be used to send POST/GET requests to FastAPI running on Heroku or locally.

GitHub workflow config

environment.yml

Continuous Integration is provided by GitHub. pylint and pytest are executed and so before the app is developed to Heroku, the checks have to pass on GitHub.

Contact

Project Link: https://github.com/MidnightSkyUniverse/census_classifier_with_fastAPI