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.
- 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.
- develop (Python): Experimental code to try out new ideas and experiments. Use Jupyter notebooks wherever you can. Naming convention:
- 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.
- 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
<dataset-desc>_<preprocessing-desc>_<training-desc>.<filetype>
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 | 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. |