Prob prog for ill posed inverse imaging problems

I’m looking for resources for pyro for a course project. I’m taking probabilistic programming with Frank Wood at the University of British Columbia.

My course project involves single particle electron cryomicroscopy, which has a detailed forward model of image formation from electron optics and protein simulation. It images microscopic things (sub nm scale). It’s a typical ill posed inverse imaging problem. Similar tomographic projection math as computerized tomography and also some other types of medical imaging.

I’ve coded up some simple examples (e.g. Gaussian unknown mean) we did in our course in pyro, and looked through some of the pyro documentation.

The observes (pyro.sample(…, obs)) will be pixelated 2D images (128x128 to 512x512 range), and there will be 10k - 1 million of them. I’d like to make sure I’m doing the panels in a way that is efficient for I/O.

For now I just want to experiment with complicated physics-y forward models, and see how good the posterior distributions are around latent variables. Try a bunch of inference approaches and see what is working in a reasonable amount of time.

Are there any tutorials for image analysis? Or github projects? Pl

If things work out nicely, my collaborators and I at SLAC national lab might incorporate more prob prog approaches into our work on 2D to 3D reconstruction using deep end to end differentiable generative models. ICCV 2021 Open Access Repository

Please solicit your proud links to related projects.

Hi @geoffwoollard, you might start with the VAE tutorial, the Attend Infer Repeat tutorial, and the Deep Kernel Learning example (these all use the MNIST dataset).

Does anyone else have examples of image analysis models in Pyro?

here are two links

the best way to look for other projects/repos is probably to browse through the titles and abstracts of papers that cite pyro