HD 143006 Imaging Tutorial Part 3

[iancze]

An overview of the HD 143006 imaging series is provided in #25

Part III is meant to be the "production-ready" version of a MPoL imaging script. The idea is to use computational resources to do hyperparameter sweeps. One example is using Ray Tune, though we might also want to suggest alternatives.

• Explore "production-ready" scripting layouts (no Jupyter notebooks). Some useful code and directory structures here.
• We have a starting point in the examples/HD143006 subdirectory. These need to be improved to be more generalizable (no hardcoded paths) similar to the examples in the Pytorch directory. We also need to document how to visualize loss, etc. with tensorboard. * Explore thorough hyperparameter testing with Ray Tune.
• Develop new priors if necessary to achieve a good image

This "tutorial" document might not actually be a .py file -> Jupytext -> ipynb like the other tutorials. Rather, it might just be a regular .rst file with text describing how one might go about running the each of the steps. If necessary, we can include a final notebook that uses the optimal parameters to visualize the best-fit image.

[jeffjennings]

Suggest adding a step demonstrating 'pre-training' to a dirty image (#206) to this tutorial.

[jeffjennings]

Suggest adding mention of use of a learning rate scheduler (#207) to this tutorial.

[jeffjennings]

Suggest adding the training diagnostic figure generated by #208 to this tutorial.

[iancze]

All sounds good. I suggest we descope Ray from the original issue, as well. Depending on how the training and crossval stuff ends up, I think we should also revise Tutorial 2, or possibly combine 2&3.

[jeffjennings]

Suggest adding the image comparison figure generated by #211 to this tutorial.

[jeffjennings]

All sounds good. I suggest we descope Ray from the original issue, as well. Depending on how the training and crossval stuff ends up, I think we should also revise Tutorial 2, or possibly combine 2&3.

Cool, I strikethrough'ed the Ray bullet. And that sounds great. Seems like a good time to come to this would be once the total flux prior is implemented and a couple more diagnostic figures I want to add are done. #142 would also be worth addressing (adding figures made by vis_histogram_fig and split_diagnostics_fig figures to the tutorials) when we address this.

There's a separate issue for this #134, but I'd also like to have a tutorial on the fit 'pipeline' (running a full optimization workflow with python -m mpol.pipeline) when that's done.