loreloc
commented
52 minutes ago Hi Tom, I took a look at your library and I found it a very cool work!
One interesting integration with cirkit could be rewriting your code such that it converts a tree-shaped PGM into our symbolic circuit representation. I think this can also be very useful to us to understand if we need to refine or extend our symbolic representation.
I also found very nice how you deal with random variables where, e.g., you can name them. Right now, we only have a scope data structure where variables are instead non-negative integers, which I guess it's too low-level as a representation and instead would benefit from your approach.
For the functional vs object-oriented design choice: I think our library benefits from a functional design as in my view it is a language and a compiler, and since it can be quite hard to implement we care about having code that we can check the correctness more easily. For instance, in our case this is done by having immutable objects that are always consistent up to initialization.
Thanks
Well met! I am a researcher at the Institute for Artificial Intelligence at the University of Bremen. My area of study is tractable probabilistic cognitive robot plans. For this purpose I have written the package probabilistic_model (https://github.com/tomsch420/probabilistic_model) which contains PGMs in networkx and PCs in networkx, jax and torch. I talked with Antonio about my progress and a potential integration with your framework. I would be very happy to do so, since we can drastically reduce the amount of duplicated work while combining our knowledge to build a better framework. My architecture is inspired by the one that Anji Liu presented (https://arxiv.org/pdf/2406.00766) but bets on sparse tensors. Currently, my work indicates that JAX has the best integration for that. While I do not have an extensive benchmark yet, I testet with ~90k parameters and 200k samples. One forward pass took approximately 50ms with JAX in that situation, while the sum layers were extremely sparse. Furthermore, my Architecture is completely object oriented. From your code I saw a lot of functional design and factories, which Python is not build for. (Also indicated by #276) I would be happy to discuss the architectures in detail and assist you in building a framework that is fit for all kinds of users.
Greetings, Tom