make apt works

[Justinezgh]

Hi @EiffL here is a notebook to test apt on a toy gaussian model:

where the goal is to learn the posterior distribution (in red) from the proposal (in green).

To validate the loss I plan to follow these steps:

With not smooth NF:

1.  Learn the posterior from the prior
  a. with nll loss
  b. with apt loss 

-> should get the same contours. 

If validated:

2.  Learn the posterior from the proposal
  a. with nll loss (weird behavior is expected)
  b. with apt loss (should get the correct contours)

-> if it works then apt loss is validated and we can test it on our smooth nf

Do the same with Smooth NF

In the notebook you can choose which experiment you want to perform through:

smooth = False
proposal = False
loss_apt = False

[Justinezgh]

And turns out that I'm stuck at the very beginning >.< I'm not able to learn the posterior distribution from the prior with a 'normal' NF

[EiffL]

Here is a working example of APT in the following simple Gaussian case:

If we don't correct for the proposal in the NPE we get this:

If we do, with an APT loss:

Notebook here: https://colab.research.google.com/drive/1CsLCftpw3K1F7ntHUveKHiBBMSG59mgE?usp=sharing

[Justinezgh]

Toy model:

Not Smooth NF

• learn posterior from prior with nll loss
  

• learn posterior from prior with apt loss
  

• learn posterior from proposal with nll loss
  

• learn posterior from proposal with apt loss
  

Smooth NF

• learn posterior from prior with nll loss
  

• learn posterior from prior with apt loss
  NaNs

• learn posterior from proposal with nll loss
  

• learn posterior from proposal with apt loss
  NaNs

Notebook here

[EiffL]

ok, so what's going on with this smooth NF? where do the NaNs occur? and what level?