Add jitter_scale parameter for initial point generation #7643
Conversation
|
|
| @@ -66,6 +66,7 @@ def make_initial_point_fns_per_chain( | |||
| model, | |||
| overrides: StartDict | Sequence[StartDict | None] | None, | |||
| jitter_rvs: set[TensorVariable] | None = None, | |||
| jitter_scale: float = 1.0, | |||
There was a problem hiding this comment.
Need to update the docstrings of the function that now accept jitter_scale
| @@ -152,6 +152,34 @@ def test_adds_jitter(self): | |||
| assert fn(0) == fn(0) | |||
| assert fn(0) != fn(1) | |||
|
|
|||
| def test_jitter_scale(self): | |||
There was a problem hiding this comment.
Can you simplify this test? If you test the default 1.0 jitter and a very large amount like 1000.0 a single draw should be enough to test the change.
There was a problem hiding this comment.
Would I need to test if the large jitters are close to +- 1000? If so, what if the single draw for the large jitter is +/- 80? I'm not sure how to test if the jitter_scale is correct without making many draws.
There was a problem hiding this comment.
The large jitter should be >> 10 in magnitude, with p=1-(10/1000)=0.99, whereas with default jitter should be smaller than 1, with p=1
|
|
||
| assert np.all((-0.05 < init_standardised) & (init_standardised < 0.05)) | ||
| assert fn_large(0)["A_log__"] > fn_default(0)["A_log__"] |
There was a problem hiding this comment.
| assert fn_large(0)["A_log__"] > fn_default(0)["A_log__"] | |
| assert fn_large(0)["A_log__"] > 10 | |
| assert fn_default(0)["A_log__"] < 1 |
There was a problem hiding this comment.
Or if 10 is too large, something smaller
ricardoV94
changed the title
Description
Add jitter_scale parameter to control the bounds of uniform jitter in initial points. This allows users to adjust the width of random initialisation, which is useful for the Multi-path Pathfinder algorithm to explore posteriors in a larger parameter space or where the posterior mean is much further away from the prior mean.
Changes:
Related Issue
Checklist
Type of change
📚 Documentation preview 📚: https://pymc--7643.org.readthedocs.build/en/7643/