ekharitonov / capstone_project_ai-enterprise

AI Enterprise Workflow - Capstone Project Submission

Capstone project submission for the IBM AI Enterprise Workflow course on Coursera.

Usage

Start application.

python run_app.py

Test application.

python run_tests.py

Predict future revenue (default is total revenue for next 30 days; add country parameter to get revenue for specific country).

curl --request POST 'http://127.0.0.1/predict?date=2018-11-20'

Build image.

docker build -t app .

Run image.

docker run \
    -it \
    --rm \
    -p 3000:80 \
    --name app \
    app

Endpoints

• POST /predict?date={date}&duration={duration}&country={country}
• POST /logs?type={type}

Marking Criteria

• Are there unit tests for the API?
  Yes, see tests/app_test.py.
• Are there unit tests for the model?
  Yes, see tests/model_test.py.
• Are there unit tests for the logging?
  Yes, see tests/log_test.py.
• Can all of the unit tests be run with a single script and do all of the unit tests pass?
  Yes, run python run_tests.py.
• Is there a mechanism to monitor performance?
  Yes, see src/monitor.py which contains a function to compute the Wasserstein distance metric.
• Was there an attempt to isolate the read/write unit tests from production models and logs?
  Yes, see src/log.py.
• Does the API work as expected? For example, can you get predictions for a specific country as well as for all countries combined?
  Yes, use curl --request POST 'http://127.0.0.1/predict?date=2018-11-20' or curl --request POST 'http://127.0.0.1/predict?date=2018-11-20&country=Australia'
• Does the data ingestion exists as a function or script to facilitate automation?
  Yes, see src/ingest.py.
• Were multiple models compared?
  Yes, an ARIMA and SARIMA model were compared. Model comparisons are in nb/results.ipynb.
• Did the EDA investigation use visualizations?
  Yes, see nb/analysis.ipynb which includes time-series, seasonal-trend decomposition, auto-correlation and partial auto-correlation plots.
• Is everything containerized within a working Docker image?
  Yes, see Dockerfile.
• Did they use a visualization to compare their model to the baseline model?
  Yes, see nb/results.ipynb where the ARIMA and SARIMA model results are compared to the actual revenue.