Hello,

I am relatively new to this language and concept, so forgive me for my naivete.

I am attempting to use the inferential framework allowed by Pyro. I am finding the parameterized distributions of PyTorch to be relatively intuitive. The model/guide framework up to this point seems to run without issue, but I believe I have a “kink”, so to speak, to the loss back-propagation (since I cannot converge my model to the expected output).

In my model/guide I have functional mapping of `a->a'`

. For effect, this map can be represented as a square, 2D matrix with `a`

in the rows and `a'`

in the columns. Each value of this matrix represents a probability of `a`

being observed as `a'`

. I am hoping to utilize this matrix `P(a|a')`

and its easily computable cousin `P(a'|a)`

to infer the input of my model given a certain output measurement further down the line.

In my efforts, I have come across the concept of PyTorch `Transformed Distributions`

, but I am uncertain of their utility in this situation. Additionally, I have looked into Categorical distributions.

Any tips/tricks/help would be very much appreciated!