Add stratified resampling; improve handling of -inf weights

[benlebrun]

This PR adds stratified resampling to smc_standard. The resampling
method can be specified by a new resampling_method argument to
smc_standard which defaults to multinomial (and thus preserves the
past behaviour).

I have also improved the handling of cases in which all particles have
-inf weight. Previously our logsumexp method would return nan and raised a warning. I've added a case to logsumexp which
returns -inf when all input weights are -inf. We also skip the
resampling step (and ess_threshold check) when all
particles are -inf.

[codecov]

Codecov Report

Attention: Patch coverage is 86.20690% with 4 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
llamppl/inference/smc_standard.py	88.46%	3 Missing ⚠️
llamppl/util.py	66.66%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!

[Copilot]

Pull Request Overview

This PR introduces stratified resampling and improves handling of cases with -∞ weights in the sequential Monte Carlo (SMC) inference code. Key changes include updating test parameters to include stratified resampling, adding a new stratified_resample function in the inference module, and propagating the new resampling_method parameter to example scripts.

Reviewed Changes

Copilot reviewed 6 out of 6 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
tests/test_examples.py	Updated test cases to use fewer particles and added stratified tests
llamppl/util.py	Improved logsumexp by converting inputs to an array and handling -∞
llamppl/inference/smc_standard.py	Added stratified_resample function and integrated the new resampling option; updated resampling condition
examples/hard_constraints.py	Updated run_example signature to accept and forward resampling_method
examples/haiku.py	Updated run_example signature to accept and forward resampling_method

Comments suppressed due to low confidence (1)

llamppl/inference/smc_standard.py:116

• [nitpick] For better performance, consider using np.random.choice with the size parameter to generate ancestor_indices in one call, rather than using a list comprehension.

ancestor_indices = [np.random.choice(range(len(particles)), p=probs) for _ in range(n_particles)]

[Copilot]

[nitpick] Consider using np.arange(N) instead of range(N) for clarity and to ensure consistent behavior with numpy operations.

Suggested change

	positions = (np.random.random(N) + range(N)) / N
	positions = (np.random.random(N) + np.arange(N)) / N