Skip to content

Latest commit

 

History

History
122 lines (99 loc) · 5.54 KB

README.md

File metadata and controls

122 lines (99 loc) · 5.54 KB

Reproducible Science

A boilerplate for reproducible and transparent science with close resemblances to the philosophy of Cookiecutter Data Science and Mario Krapp's (cookiecutter)[https://github.com/mkrapp/cookiecutter-reproducible-science]: A logical, reasonably standardized, but flexible project structure for doing and sharing data science work.

This template is designed to document the creation of a dataset/presentations/papers. It has been refined over several research projects at the (UChicago Justice Project)[https://github.com/uchicago-justice-project].

An idealized data science workflow:

flowchart TD
    A("fa:fa-file data/raw/") -->|fa:fa-terminal clean.py| B("fa:fa-file data/interim/v1")
    B --> |fa:fa-terminal more_cleaning.py| C("fa:fa-file data/final/v2")
    C --> D[fa:fa-archive academic talk]
    C --> E[fa:fa-book academic paper]
Loading

Project Structure

.
├── AUTHORS.md
├── LICENSE
├── README.md
├── data                    <- (not tracked by git. maybe tracked by DVC)
│   ├── raw                 <- The original, immutable data dumps from primary or third-party sources.
│   ├── interim             <- Intermediate data that has been transformed.
│   ├── final               <- The final, canonical data sets for modeling.
│   └── replication         <- Final data that was used in a paper, talk, etc.
|       └── my-paper        <- One folder per task/presentation
├── docs                    <- Documentation, e.g. data dictionaries, memos, project notes
├── notebooks               <- Interactive python or R notebooks
│   ├── exploratory         <- One-offs (one script per task)
│   └── replication         <- Report-generating scripts that were used in a paper, talk, etc. (one script per presentation)
├── pipeline                <- The actual data ETL process (a NUMBERED mix of scripts and notebooks)
├── reports                 <- For all non-data project outputs
|   ├── exploratory         <- One-offs
|   |   └── my-task         <- One folder per task/presentation
|   |       ├── figures     <- Visualizations
|   |       ├── tables      <- Descriptive tables
|   |       └── regressions <- Model outputs
│   └── replication         <- Outputs that were presented in a talk, paper, etc. (promote exploratory subfolders into here)
├── setup.py                <- Allows importing python files from src into notebooks
└── src                     <- Refactored helper code
    ├── data                <- Resusable data ETL steps
    ├── stats               <- Source code for modeling and statistics
    ├── viz                 <- Scripts for visualisation of your results, e.g., matplotlib, ggplot2 related.
    └── util                <- Any helper scripts go here

Setup

Install cookiecutter command line:

pip install cookiecutter

For the error: "Unable to load extension: No module named 'jinja2_time'"

pip install jinja2_time

To start a new science project:

cookiecutter gh:eric-mc2/cookiecutter-reproducible-science

Create local python package from src to allow relative imports in notebooks (optional):

cd <repo-name>; pip install -e

Usage / Recommendations

Don't painstakingly refactor your project to fit into this mold, but do use the parts of this template that make sense in the context of your research goals. This is a loosely-opinionated structure to help with some of the following goals:

  • document the streamlined/idealized data generation process leading to the finalized dataset
  • document the actual data generation process of an in-progress dataset
  • document validation steps that motivated further data processing steps
  • store data versions to triage issues and affected analyses

The opinionated workflow looks like this:

  1. Pull your data into data/raw.
  2. Apply transformations. a. Transformed data goes in data/interim. b. All transformations increment the data version number, e.g. (census_v1.csv). c. Keep previous data versions! Storage is cheap! e. The highest-version files that are part of the "final product" go in data/final. f. Whenever you revise the "final" data, just move the file to interim, and add the new incremented version to final.
  3. Document all transformations in pipeline/ a. If the process was manual, still document it as a comment-only script. b. Number all pipeline files in the order they should run.
  4. Promote all data, outputs, and scripts related to "published" work into replication.

A more realistic research data science workflow:

flowchart TD
    RAW("fa:fa-file data/raw")
    RAW --> EDA[/fa:fa-terminal notebooks/exploratory/eda.py/]
    EDA --> EXP[fa:fa-archive reports/exploratory/first-chart.png]
    CLEAN[/fa:fa-terminal pipeline/cleaning stages.py/]
    RAW -->CLEAN
    CLEAN --> INT("fa:fa-file data/interim/v++")
    INT --> CLEAN
    REG1[/fa:fa-terminal notebooks/exploratory/modeling/test/]
    INT --> REG1
    REG1 --> TALK[fa:fa-archive reports/exploratory/test-regression.tex]
    MORE(fa:fa-file data/raw/covariates)
    MORE --> MERGE[/fa:fa-terminal pipeline/merge.py/]
    INT --> MERGE
    MERGE --> FIN("fa:fa-file data/final")
    REG2[/fa:fa-terminal notebooks/replication/asr-paper/]
    FIN --> REG2
    REG2 --> PAPER[fa:fa-book academic paper]
Loading

License

This project is licensed under the terms of the BSD License