Implement ML tasks

[Justinezgh]

Tasks (from https://arxiv.org/pdf/2101.04653.pdf):

• Gaussian 2D (✔️)

• Gaussian mixture(✔️)

• Two Moons (✔️ but not 100% sure )

• Lotka-Volterra (⌛)

[Justinezgh]

notebook : here

[Justinezgh]

For the 2 moons, should I use dist.TransformedDistribution(distribution, transform) from numpyro to get the simulation x ?

[EiffL]

Hum you don't necessarily need to use numpyro for anything here. You can do everything in tfp

[Justinezgh]

I'm not sure to understand, can we get p(theta|x,z) with tfp ?

[EiffL]

Hummmm I'm not sure I understand either ^^ numpyro doesn't give you p(theta|x,z) either?

[Justinezgh]

ok I'm lost ^^

[Justinezgh]

Lokta-Volterra: notebook

I still have a pb with vmap() or Independent() because when I increase the time or the batch_size the score increases too. But observations seems ok

[EiffL]

Cool cool cool :-D

[Justinezgh]

from: https://arxiv.org/abs/1905.07488

[Justinezgh]

I learned Lokta Volterra posterior with 2 different compressors: one trained with MSE and one with VMIM.

• MSE: notebook

• VMIM: notebook

[EiffL]

That's awesome! Very nice! So the regression compression doesn't get any better than that no matter what you do?

[EiffL]

I'll just ping @dlanzieri about this because this is exactly the sort of thing we are interested in regarding this compression question

[Justinezgh]

you can probably do better ^^ but the regression seems not that bad: (orange points are jnp.abs(reg[:,j]-p[:,j])>0.05 with p the parameters and reg the prediction)