Sparse Probabilistic Circuits

[bhunecke]

This PR adds pruning and growing functionality to the probabilistic circuits implementation. See Dang et al. Sparse Probabilistic Circuits via Pruning and Growing for more details.

Try prune and grow separately using the example scripts in the respective folder.
The jpt_pruning_growing_demo.py and jpt_pruning_growing_mnist.py demo scripts show how pruning and growing can be applied on a learned JPT.

[tomsch420]

Proivde unittests and a notebook in the docs that showcases pruning and growing (refactor your examples to a markdown notebook, see the doc notebooks for refrence)

[tomsch420]

This is way nicer than before but not fully finished yet. The comments tell you what to do. Also read this: https://testing.googleblog.com/2017/11/obsessed-with-primitives.html and replace the dict types by proper registries.

[Copilot]

Pull Request Overview

This PR implements pruning and growing functionality for probabilistic circuits based on the Sparse Probabilistic Circuits paper by Dang et al. The implementation enables structural optimization of circuits by removing less important edges (pruning) and adding new components with noise (growing).

• Adds flow analysis capability to compute edge importance based on data flows
• Implements pruning method to remove low-importance edges based on flow analysis
• Implements growing method to duplicate circuit structure with noise injection

Reviewed Changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 7 comments.

Show a summary per file

File	Description
src/probabilistic_model/probabilistic_circuit/rx/probabilistic_circuit.py	Adds prune() and grow() methods to ProbabilisticCircuit class
src/probabilistic_model/probabilistic_circuit/rx/flow_analyzer.py	New CircuitFlowAnalyzer class for computing edge flows through circuits
doc/references.bib	Adds citation for the sparse probabilistic circuits paper
doc/pruning_growing.md	Comprehensive tutorial demonstrating pruning and growing techniques
doc/_toc.yml	Adds the new tutorial to the documentation table of contents

---

_{Tip: Customize your code reviews with copilot-instructions.md. Create the file or learn how to get started.}

[Copilot]

[nitpick] The prune method modifies the circuit in-place and returns self, but this could be confusing since users might expect it to return a new pruned circuit. Consider either making this operation purely in-place (return None) or making it return a new circuit instance to avoid ambiguity.

[Copilot]

[nitpick] Similar to the prune method, the grow method modifies the circuit in-place and returns self, which could be confusing. Consider making the API consistent by either returning None for in-place operations or returning new instances.

[Copilot]

These identity checks using 'any()' with generator expressions are inefficient for large circuits. Since old2new is a dictionary, use 'parent in old2new' and 'parent in old2new.values()' instead, or better yet, use sets to track which nodes belong to which group.

[Copilot]

Import statements should be grouped and sorted. Standard library imports (collections, typing) should come first, followed by third-party imports (numpy, tqdm), then local imports. Consider using 'from tqdm import tqdm' for cleaner code.

[Copilot]

Hard-coded progress bar description should be configurable or removed for library code. Consider adding a 'show_progress' parameter to allow users to control progress bar display.

Suggested change

	def compute_flows(self, dataset: np.ndarray) -> Dict[Tuple['Unit', 'Unit'], float]:
	"""
	Compute the flow of information through the circuit for a given dataset.
	:param dataset: The input dataset.
	:return: Dictionary mapping edge tuples to their flow values.
	"""
	edge_flows = defaultdict(float)
	for x in tqdm.tqdm(dataset, desc="Computing circuit flows"):
	def compute_flows(self, dataset: np.ndarray, show_progress: bool = False) -> Dict[Tuple['Unit', 'Unit'], float]:
	"""
	Compute the flow of information through the circuit for a given dataset.
	:param dataset: The input dataset.
	:param show_progress: If True, display a progress bar during computation. Default is False.
	:return: Dictionary mapping edge tuples to their flow values.
	"""
	edge_flows = defaultdict(float)
	iterator = tqdm.tqdm(dataset, desc="Computing circuit flows") if show_progress else dataset
	for x in iterator:

[Copilot]

Division by zero protection is insufficient. While checking parent._flow_likelihood > 0, there's no check for node._flow_likelihood existence before the division. This could cause issues if node doesn't have the attribute but parent does.

Suggested change

	if hasattr(parent, '_flow_likelihood') and parent._flow_likelihood > 0:
	if hasattr(node, '_flow_likelihood'):
	contribution = node._flow_likelihood / parent._flow_likelihood
	node_flows[node] += contribution * node_flows[parent]
	if hasattr(parent, '_flow_likelihood') and parent._flow_likelihood > 0 and hasattr(node, '_flow_likelihood'):
	contribution = node._flow_likelihood / parent._flow_likelihood
	node_flows[node] += contribution * node_flows[parent]

[Copilot]

Potential division by zero if node._flow_likelihood is zero or very close to zero. Add a check to ensure node._flow_likelihood > 0 before performing the division.