PyMC/PyTensor Implementation of Pathfinder VI #387
Conversation
`fit_pathfinder` - Edited `fit_pathfinder` to produce `pathfinder_state`, `pathfinder_info`, `pathfinder_samples` and `pathfinder_idata` for closer examination of the outputs. - Changed the `num_samples` argument name to `num_draws` to avoid `TypeError` got multiple values for keyword argument 'num_samples'. - Initial points are automatically set to jitter as jitter is required for pathfinder. Extras - New function 'get_jaxified_logp_ravel_inputs' to simplify previous code structure in fit_pathfinder. Tests - Added extra test for pathfinder to test pathfinder_info variables and pathfinder_idata are consistent for a given random seed.
Add a new PyMC-based implementation of Pathfinder VI that uses PyTensor operations which provides support for both PyMC and BlackJAX backends in fit_pathfinder.
- Implemented in to support running multiple Pathfinder instances in parallel. - Implemented function in for Pareto Smoothed Importance Resampling (PSIR). - Moved relevant pathfinder files into the directory. - Updated tests to reflect changes in the Pathfinder implementation and added tests for new functionalities.
…race data to InferenceData
… for bfgs_sample
aphc14
force-pushed
the
pathfinder_w_pytensor_symbolic
branch
from
9bfc48c
to
ef2956f
Compare
aphc14
changed the title
|
This version runs much faster than #386, but the codes are messier due to the numerous pytensor symbolic variables created for the compiled pytensor functions (see the lines of code between |
| g: np.ndarray | ||
|
|
||
|
|
||
| class LBFGSHistoryManager: |
There was a problem hiding this comment.
Cleaner to use a data class? Don't know.
There was a problem hiding this comment.
yep, I agree. dataclass now added
Summaryh of changes: - Remove multiprocessing code in favour of reusing compiled for each path - takes only random_seed as argument for each path - Compute graph significantly smaller by using pure pytensor op and symoblic variables - Added LBFGSOp to compile with pytensor.function - Cleaned up codes using pytensor variables
…and . - Corrected the dimensions in comments for matrices Q and R in the function. - Uumerical stability in the calculation by changing from to .
pymc_experimental/inference/fit.py
Outdated
| @@ -31,11 +31,13 @@ def fit(method, **kwargs): | |||
| arviz.InferenceData | |||
| """ | |||
| if method == "pathfinder": | |||
| # TODO: Remove this once we have a pure PyMC implementation | |||
There was a problem hiding this comment.
This PR will provide that, no?
There was a problem hiding this comment.
the latest commit addresses this
Fixed incorrect and inconsistent posterior approximations in the Pathfinder VI algorithm by: 1. Adding missing parentheses in the phi calculation to ensure proper order of operations in matrix multiplications 2. Changing the sign in mu calculation from 'x +' to 'x -' to match Stan's implementation (which differs from the original paper) The resulting changes now make the posterior approximations more reliable.
Implements both sparse and dense BFGS sampling approaches for Pathfinder VI: - Adds bfgs_sample_dense for cases where 2*maxcor >= num_params. - Moved existing and computations to bfgs_sample_sparse, making the sparse use cases more explicit. Other changes: - Sets default maxcor=5 instead of dynamic sizing based on parameters Dense approximations are recommended when the target distribution has higher dependencies among the parameters.
Bigger changes: - Made pmx.fit compatible with method='pathfinder' - Remove JAX dependency when inference_backend='pymc' to support Windows users - Improve runtime performance by setting trust_input=True for compiled functions Minor changes: - Change default num_paths from 1 to 4 for stable and reliable approximations - Change LBFGS code using dataclasses - Update tests to handle both PyMC and BlackJAX backends
- Add LBFGSInitFailed exception for failed LBFGS initialisation - Skip failed paths in multipath_pathfinder and track number of failures - Handle NaN values from Cholesky decompsition in bfgs_sample - Add checks for numericl stabilty in matrix operations Slight performance improvements: - Set allow_gc=False in scan ops - Use FAST_RUN mode consistently
Major: - Added progress bar support. Minor - Added exception for non-finite log prob values - Removed . - Allowed maxcor argument to be None, and dynamically set based on the number of model parameters. - Improved logging to inform users about failed paths and lbfgs initialisation.
|
Need to make an important change related to how important sampling is done. Based on some tests, for trickier posteriors, psir (Pareto smoothed importance resampling) tends to cause many large peaks. In contrast to the reference posterior (what you’d get using NUTS), it doesn’t have such peaks. Turning off resampling, you'd get psis instead, and the final posterior better resembles NUTS, so you don't get the weird peaks behaviour. But this would differ from the original paper, which uses psir. Since the choice of importance sampling (IS) can have a big impact on the final posterior, and there are several IS methods, I plan to use a class variable that controls how IS is done based on the user inputs. I'm thinking of making psis (and not psir) the default IS behaviour as the safest and most generally reliable option. Shouldn't take long to fix. |
Another version to draft PR #386 which uses more of PyTensor's symbolic variables and compiling functions.
Questions for Review