demo-github-action

Predict satellite data based on L3 data (including missing data) with a UNet

Figure: SST in degree Celsius for the first (independent) test data.

Input:

• Sea surface temperature (SST, Level 3, i.e. with missing data) from the 7 past days

Output:

• SST prediction of the next with without clouds
• Expected standard deviation error of the SST prediction

Model:

• Standard UNet (https://doi.org/10.1007/978-3-319-24574-4_28)

Loss function:

• Mean squared error or
• DINCAE loss function, likehood of the observation for a Gaussian pdf with a given mean and standard deviation (see https://doi.org/10.5194/gmd-13-1609-2020 for details)

Github action

• Why continuous integration testing? Detect failures early.
• Test your code on every commit, pull request
• Possibly on multiple platforms (Ubuntu, Mac OS,...) and different versions of julia/python/...
• All these combinations represent the test matrix
• However, even if GitHub action is free for open source projects, consider the environmental impact and avoid unnecessary combinations of OS/version/...

Recommentations

• Automate the installation of all dependencies by declaring them in the requirements.txt or pyproject.toml files
• First make sure that you can instanticate the project from a clean environement and run all tests locally
• There is no GPU on github actions
• Tests should be short
• Badge for your README.md:

[![Build Status](https://github.com/neccton-algo/demo-github-action/workflows/CI/badge.svg)](https://github.com/neccton-algo/demo-github-action/actions)

• Notebooks can be run and tested with nbconvert
• There is an environement variable CI which is set to true on github action to adapt the code path to be tested.
• For python: consider testing frameworks like py.test, Hypothesis, tox, ...
• For julia: write test in the test/runtests.jl which gets executed with Pkg.test().
• Monitor code coverage with e.g. codecov.io.