Moving to L0 package for HardConcrete regularization implementation

[juaristi22]

Fix #84

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Preview	Comments	Updated (UTC)
microcalibrate	Ready	Preview	Comment	Aug 18, 2025 11:12am

[juaristi22]

All existing tests (including a specific regularization test) passed though maybe we want to add some more comprehensive ones. Any specific gate behavior it is crucial to capture @baogorek ?

[baogorek]

All good. I was just a bit confused because I thought we had moved to a claude-code refactored version of the code at https://github.com/AMLab-Amsterdam/L0_regularization. Now I see that it is a fork of that codebase, but that we have brought in our own HardConcrete class, so by importing that class, we're bypassing the functionality from the external github package.

I think that's fine for now, as this PR still contributes modularity to this repo. I still have doubts about our L0 implementation (which I vibe-coded), so I will spend some time there. FYI @MaxGhenis

[MaxGhenis]

I don't follow Ben - isn't this moving from your version of L0 to the version I vibe-coded in the L0 package?

[baogorek]

@MaxGhenis , what led me to believe that was noticing that the interface to the procedure is exactly identical in this PR:

Then, here's what Claude Code said about it:

  PolicyEngine's l0-python package is a hybrid:                                                                                                                             
                                                                                                                                                                            
  1. HardConcrete class: This is YOUR original implementation from src/microcalibrate/utils/l0.py                                                                           
    - Same exact interface: HardConcrete(input_dim, output_dim=None, temperature=0.5, stretch=0.1, init_mean=0.5)                                                           
    - Same get_penalty() method implementation                                                                                                                              
    - Just added more documentation                                                                                                                                         
  2. New additions (not in your original code):                                                                                                                             
    - layers.py with L0Linear, L0Conv2d, etc.                                                                                                                               
    - gates.py with various gate types                                                                                                                                      
    - penalties.py with penalty computation utilities                                 
  3. Not a direct AMLab fork:                                                                                                                                               
    - AMLab has L0Dense with different parameters: droprate_init, lamba, local_rep                                                                                          
    - PolicyEngine has L0Linear with different parameters: init_sparsity, use_l2                                                                                            
    - Different interfaces and naming conventions            

Yeah, I just checked. There's no HardConcrete class in the AMLab-Amsterdam repo. I'm guessing Claude Code knew about it when it was doing the cleanup?

Not surprisingly (since it's parent LLM made it), Claude loves our HardConcrete class and thinks we should keep using it.