[Baselines] Basic diffusion model generative baseline

[ConnorStoneAstro]

Basic diffusion model generative baseline

Train a diffusion model as a generative model for one of the datasets in AstroPile

Contacts: @ConnorStoneAstro Participants:

Goals and deliverable

Set up framework to train a diffusion model on an AstroPile formatted dataset. This would be for a single mode.

Resources needed

Going to pick a low dim dataset for the hack. Framework should be usable on any dataset

Detailed description

[ConnorStoneAstro]

Manage to load data

[ConnorStoneAstro]

Using my laptop to train a model. Using all my CPU

[ConnorStoneAstro]

And now I have samples!!! They are bad because I only trained for 5min on CPU

[ConnorStoneAstro]

Had to hack the score_models package to use an IterableDataset which doesn't have a len value. It's very trivial changes so I will make a PR for score_models

[ConnorStoneAstro]

Also, had to enforce that the spectrum was a power of 2. This is because the default score model is a (1D) U-net, this can be up to the user if you want something fancy.