Remove _Args and rework model-level interfaces

april-tools / cirkit

a python framework to build, learn and reason about probabilistic circuits and tensor networks

https://cirkit-docs.readthedocs.io/en/latest/

GNU General Public License v3.0

71 stars 1 forks source link

Closed loreloc closed 1 year ago

loreloc commented 1 year ago

Removed that ugly _Args class.
Removed all reference to "k" and use num_*_units instead.
Reworked the public interfaces of model:
- Added layer_kwargs for specifying layer-agnostic hyper-parameters (e.g., rank for CPLayer).
- Rename the model to TensorizedPC.
Made input layer and inner layers separate attributes.
Overall code should be more readable (e.g., no mysterious _k_gen() function) and slightly more documented.

Closes #72 .

loreloc commented 1 year ago

I have rebased the changes about generalizing the log-sum-exp trick from main, and then broke down this in smaller commits.

lkct commented 1 year ago

This unfortunately required major refactoring to disentangle hyper-parameters and also concepts/nomenclature

BTW just "removing _Args" is actually very easy. This PR is actually doing much more than that.