Python-related procedures

In this section, we will show you a few things related specifically to running code reproducibly with Python and Jupyter notebooks. . For more general debugging tips for Python and other computer languages, see our wiki.

Best practice to reproduce a Python paper (Python environments)

You should create a Python environment that is dedicated to the project. See Anaconda instructions as one possible method, venv as another one, though others exist.

Here’s venv version in a nutshell (full guide)

• Ensure venv exists:

pip3 install pyenv

• Create a new environment

python3 -m venv /path/to/new/virtual/environment

or if using relative paths

python3 -m venv env

which will create /path/to/new/virtual/environment or (relative to your current working directory) env. That directory will now contain all of your project-related Python packages.

To activate:

source env/bin/activate

On Windows Bash (depends on install)

python -m venv env
source env/Scripts/activate

To deactivate:

deactivate

Making Python code dynamic

In general, Python code should not have hard-coded paths. Python programs are aware of their own location, and other directories should be relative to that. However, some authors may still follow (econ-specific) norms, and hard-code paths. Similar to our approach for Stata, R, and MATLAB, you can use the rootdir principle to make the code more portable/dynamic.

In that case, do the following:

Say the author has code like

xlsread('C:\Users\me\submission\AEJMacro\yesterday\data.xlsx')

At the top of the program, add the following lines:

import os
rootdir = os.path.realpath(__file__)

Then, wherever the hard-coded path appears, replace it with:

xlsread(os.path.join(rootdir,'data.xlsx'))

Alternatively, if the author uses a change of working directory, then

import os
os.chdir(os.path.dirname(os.path.abspath(__file__)))

would change the working directory to the location of the file being run.

Installing packages

Sometimes, authors will list the packages they used. There are a few options:

They provide a requirements.txt file

If the authors provide a requirements.txt file, you can install all the packages at once. From an appropriate terminal, run:

pip install -r requirements.txt

They provide a environment.yml file

If the authors provide an environment.yml file, they used Conda as the Python system manager. From an appropriate terminal, run:

conda env create -f environment.yml

They provide a list of packages

If the authors provide a list of packages, the easiest way is to create a simple requirements.txt file with a text editor (e.g., VS Code), then proceed as with the first option.

Using Anaconda Package manager

This is known to work on CISER (CCSS-Classic).

This may not be the way it works on CCSS-Cloud.

Needs an update

If using the default “Jupyter” link in the Start Menu, the working directory won’t be right. Assuming that you have set your Workspace to L:\Documents\Workspace, the following will create a Jupyter Notebook in the right location (thanks to Louis Liu for creating this Howto)

Search “anaconda prompt” from the start menu. right click on the app when it appears and pin it to the taskbar.

Right click on anaconda prompt in the taskbar (looks like a black window, similar to command line or terminal). Right click on “anaconda powershell prompt” in the tasks menu that pops up, and then properties.

In the properties window, go to the shortcut tab and change the “Start in:” field to U:\Documents\Workspace or whichever directory you keep your bitbucket repos in. Click apply.

Next, click on the anaconda prompt shortcut in the taskbar. When anaconda prompt opens, enter the command “Jupyter notebook”

Conda on BioHPC

If a replication package uses conda for package management, rather than pip, follow instructions at BioHPC on how to install miniconda in your home directory, then add the line

source $HOME/miniconda3/bin/activate

at an appropriate location in the code (for instance, replacing module load conda).

Running Jupyter Notebooks

Manually

In order to run Jupyter notebooks, do the following, once you have opened the Jupyter Notebook or Jupyter Lab interface:

• Navigate to the directory where the notebook is located

• Open the notebook

• Clear all the cells: Cell -> All Output -> Clear

• Run all the cells: Cell -> Run All

• Save the notebook: File -> Save and Checkpoint

From the command line

You should also be able to do the following from the command line:

Linux/macOS

If you have a LaTeX installation, you can convert the notebook to a PDF using the following commands:

# requires a latex installation
pip install nbconvert
jupyter nbconvert --to notebook --execute mynotebook.ipynb
jupyter nbconvert --to pdf                mynotebook.ipynb

Alternatively, you can convert the notebook to a PDF more closely resembling the HTML view using the following command:

pip install nbconvert
pip install pyppeteer
jupyter nbconvert --to notebook --execute mynotebook.ipynb
jupyter nbconvert --to webpdf --allow-chromium-download mynotebook.ipynb

Windows

Not sure yet. Needs augmentation.

Document the packages YOU used

If you had to iteratively install packages, you should run the following command at the end of your whole process:

pip freeze > requirements.txt

and add that output (the contents of the requirements.txt file) to the repository, and to an appendix in the report. Take care to not overwrite author-provided requirements.txt files.