/udacity_mlops_p4

Primary LanguagePython

Udacity MLOps Nanodegree Final Project

Note to reviewers: see all required generated files in generated_files folder.

Key Assumptions

  • New data files (csv) are deposited into a specified folder.
  • All files in the specified folder should be combined into a single dataset.
  • Feature columns are lastmonth_activity, lastyear_activity, number_of_employees
  • Features are numerical.
  • Label column to predict is exited.
  • Labels are binary.

ML Pipeline

Project Configuration

Using config.json the following items are set:

  • input_folder_path contains all files to be used in current dataset
  • output_folder_path a copy of the current dataset will be written here
  • test_data_path a witheld dataset in csv format for model testing.
  • output_model_path a copy of the current model will be written here
  • prod_deployment_path the model stored here is served by a web API
  • db_path path to a sqlite database

Data Ingestion

  1. Locate all files in data folder.
  2. Compile data in all files into single dataset.
  3. Deduplicate.

Model Training

  • sklearn.LogisticRegression

Diagnostics

The following diagnostic items are tracked for each dataset.

  • Column-wise mean, median, and standard deviation.
  • Column-wise percent of missing values.
  • Time required to ingest dataset.
  • Time required to train on the dataset
  • Model f1 score.
  • All pip packages installed, current and latest-available versions

Deployment

For serving of the model and associated metrics, a number of files are written to a deployment folder specified in the config.

  • trainedmodel.pkl the trained model
  • dataset_summary.csv summary statistics of the dataset used during training
  • ingestedfiles.txt list of the source files used for training
  • latestscore.txt f1 score as well as timing information
  • packages.csv pip packages installed vs latest-available version generated during training

Pipeline Automation

Automation is accomplished with a cronjob running fullprocess.py at regular intervals. This works as follows:

project pipeline

Serving

Deployed files are served on an API implemented with Flask in app.py. Tests for this API are provided in apicalls.py.

*Note one limitation in this implementation as specified by the project rubric is the API can only provide predictions for files residing on the file system of the web app.