ML Project Template

This repository contains a template project that can be easily adapted for all kinds of Machine Learning tasks. Typically, solving such task entails two main phases, research and production with very different focuses. The template intends to faciliatate work on ML projects by guiding practitioners to adopt some best practices.

research: exploratory data analyses, model prototyping and experiments are dumped here in a structured way

production: distilled utils lib, training job and inference service are implemented here

It is recommended to simply clone this repo and customize it to the specific use-case at hand.

---

Repository Structure

• research: Scripts and Notebooks for experimentation.
  • develop (Python): Experimental code to try out new ideas and experiments. Use Jupyter notebooks wherever you can. Naming convention: YYYY-MM-DD_userid_short-description. If you cannot use a notebook and have multiple scripts/files for an experiment, create a folder with the same naming convention. Each file should be handled by one person only.
  • deliver (Python): Refactored notebooks that contain valuable insights or results (e.g. visualizations, training runs). Notebooks should be refactored, documented, contain outputs, and use the following naming schema: YYYY-MM-DD_short-description. Notebooks in deliver should not be changed or rerun. If you want to rerun a deliver Notebook, please duplicate it into the develop folder.
  • templates (Python): Refactored Notebooks that are reusable for a specific task (e.g. model training, data exploration). Notebooks should be refactored, documented, not contain any output, and use the following naming schema: short-description. If you like to make use of a template Notebook, duplicate the notebook into develop folder.
• production: The production-ready solution(s) composed of libraries, services, and jobs.
  • python-utils-lib (Python): Utility functions that are distilled from the research phase and used across multiple scripts. Should only contain refactored and tested Python scripts/modules. Installable via pip.
  • training-job (Python/Docker): Combines required data exports, preprocessing and training scripts into a Docker container. This makes results reproducible and the production model retrainable in any ennvironment.
  • inference-service (Python/Docker): Docker container that provides the final model prediction capabilities via a REST API.

Naming Conventions

Code Artifacts

• develop notebooks/scripts: YYYY-MM-DD_userid_short-description
• deliver notebooks/scripts: YYYY-MM-DD_short-description
• template notebooks/scripts: short-description
• services: -service suffix
• jobs: -job suffix
• libraries: -lib suffix

Files

<dataset-desc>_<preprocessing-desc>_<training-desc>.<filetype>

Examples:

• blogs-metadata.csv
• blogs-metadata_cl-rs_ft-vec.vectors
• categories2blogs_cl-rs-lm_tfidf-lsvm.model.zip
• categories2blogs-questions_cl-rs-lm_tfidf-lsvm.model.zip

Name Identifier Descriptions:

Name	Description
Dataset Identifiers:
categories2blogs	Dataset containing blogs with the text content, blogs item URI, and connected primary tags.
blogs-metadata	Dataset containing all blogs and related metadata (properties).
Preprocessing Identifiers:
cl	Default text cleaning (lowercasing, regex cleaning).
rs	Remove Stopwords.
lm	Text lemmatization.
Training Identifiers:
ft-vec	Text vectorizer using Fasttext.
tfidf	Text vectorizer using TFIDF.
lsvm	Classifier using linear SVM.
Filetype Identifiers:
.model	Model file.
.vectors	Binary vectors file.