Example is incomplete

[tsalo]

It looks like the example doesn't actually run any of the denoising methods: https://paquiteau.github.io/patch-denoising/auto_examples/example_experimental_data.html#sphx-glr-auto-examples-example-experimental-data-py

[paquiteau]

Hi again,

Yes sadly I hadn't that much time to dedicate to patch-denoise recently, and creating an example that would be meaningfull and runnable on CI was more challenging that I thought. I am open to any idea on how to do that properly.

[tsalo]

What about running the example locally and pushing the pre-generated version instead of trying to run it on RTD? You could just re-run the test before new releases.

[paquiteau]

Well, in this case, it could be harder to reproduce for an end-user (such example would require good hardware, and/or a (very) long coffee break).

Also, I am not sure on if/how I could share example data (NORDIC/patch-denoise happens early (earliest ?) in the fMRI pipeline, so sharing raw DICOMs/nifti that have not been preprocessed (defaced notably) may also be a concern. A potential solution could be to do something based on phantoms (using [brainweb-dl](https://github.com/paquiteau/brainweb-dl, for instance).

[tsalo]

Some operations just aren't amenable to a quick example. Take something like NiMARE (a meta-analysis package I used to maintain). The examples page contains a mix of light-weight examples that are run by RTD/GitHub Actions (e.g., an overview of coordinate-based meta-analysis methods) and ones that work, but can't be run on RTD/GitHub Actions (e.g., downloading the entirety of Neurosynth or NeuroQuery). Keeping the latter set up-to-date can be annoying, but it's definitely doable.

As for test data, I've got a dataset on OpenNeuro with magnitude+phase reconstruction and 3 noise volumes (https://openneuro.org/datasets/ds005250). It's multi-echo, but you could just use the middle echo from one run. I expect a lot more compatible datasets will be uploaded in the future, given how popular the WashU/UMN 5-echo, complex-valued protocol (i.e., what's in that dataset) is becoming.