autograde is a tool for testing Jupyter notebooks. Its features include execution of notebooks (optionally isolated via docker/podman) with consecutive unit testing of the final notebook state. On top of that, an audit mode allows for refining results (e.g. grading plots by hand). Eventually, autograde can summarize these results in human and machine readable formats.

Before installing autograde and in case you want to use it with a container backend, ensure docker or podman is available on your system. We recommend podman as it runs rootless.

Now, in order to install autograde, run pip install jupyter-autograde. Alternatively, you can install autograde from source by cloning this repository and runing pip install -e . within it (or pip install -e .[develop] if you're a developer).

Eventually, build the respective container image: python -m autograde build. Note: in order to build a container image, autograde must not be installed via PyPI but from source code!

autograde comes with some example files located in the demo/ subdirectory that we will use for now to illustrate the workflow. Run:

python -m autograde test demo/test.py demo/notebook.ipynb --target /tmp --context demo/context

What happened? Let's first have a look at the arguments of autograde:

• demo/test.py a script with test cases we want apply
• demo/notebook.ipynb is the a notebook to be tested (here you may also specify a directory to be recursively searched for notebooks)
• The optional flag --target tells autograde where to store results, /tmp in our case, and the current working directory by default.
• The optional flag --context specifies a directory that is mounted into the sandbox and may contain arbitrary files or subdirectories. This is useful when the notebook expects some external files to be present such as data sets.

The output is a compressed archive that is named something like results_[Lastname1,Lastname2,...]_XXXXXXXX.zip and which has the following contents:

• artifacts/: directory with all files that where created or modified by the tested notebook as well as rendered matplotlib plots.
• code.py: code extracted from the notebook including stdout/stderr as comments
• notebook.ipynb: an identical copy of the tested notebook
• test_restults.json: test results

The report sub command creates human readable HTML reports from test results:

python -m autograde report path/to/result(s)

The respective report is added to the results archive inplace.

Results from multiple test runs can be merged via the patch sub command:

python -m autograde patch path/to/result(s) /path/to/patch/result(s)

In a typical scenario, test cases are not just applied to one notebook but many at a time. Therefore, autograde comes with a summary feature, that aggregates results, shows you a score distribution and has some very basic fraud detection. To create a summary, simply run:

python -m autograde summary path/to/results

Two new files will appear in the result directory:

• summary.csv: aggregated results
• summary.html: human readable summary report

To get an overview of all available commands and their usage, run

python -m autograde [sub command] --help