WIP: upd Monte-Carlo integration (vegasflow package) #407
Conversation
for more information, see https://pre-commit.ci
for more information, see https://pre-commit.ci
…o mc_vegasflow
mensigo
changed the title
zfit/core/integration.py
Outdated
| events_limit = vf_options.get("events_limit", int(1e6)) | ||
| list_devices = vf_options.get("list_devices") | ||
| compilable = vf_options.get("compilable", True) | ||
| verbose = vf_options.get("verbose", False) # get_verbosity() >= 0 (?) |
There was a problem hiding this comment.
Yep, we can use the global zfit settings. What are the different verbosity levels in vegasflow? zfit goes from 0 to 10
There was a problem hiding this comment.
I think, it's better to use vf's verbosity levels for convenience (4 levels) – added to docstring. However, it can be sync with zfit's levels in the future.
There was a problem hiding this comment.
The problem is that it's an additional effort to get this right. But it can also be called from outside, yeah
| @@ -280,6 +312,60 @@ def test_mc_integration(chunksize, limits): | |||
| ).numpy() == pytest.approx(integral3, rel=0.03) | |||
|
|
|||
|
|
|||
| def test_mc_vf_integration(): | |||
| # simple examples | |||
| num_integral = zintegrate.mc_vf_integrate( | |||
There was a problem hiding this comment.
There comes a crucial thing: can this run with a z.function wrapped? Because in the PDF, it usually will be.
You can also swap the default integration methods for PDFs and check if the tests still run, do they?
There was a problem hiding this comment.
I tried wrapped version and faced some issues. The thing is, vf's run_integration is not intended to run under tf.function – there are some python control flow instructions (like "if" in print_iteration and numpy math ops when computing results from iterations) and tf.Variable creation (which causes "ValueError: tf.function-decorated function tried to create variables on non-first call"). I tried to use tf.py_function, but it did not help (anyway, i think it's not a relevant option in case of graph-like z.function application). I keep on thinking how to handle this issue.
There was a problem hiding this comment.
Hm, yep. that's a bit of a problem. One thing you can try is to run it in a tf.init_scope or just make an issue with VF/ask them. That would be good to be solved somehow
| func5_sigma_i = zfit.Parameter(f"sigma_{i}", func5_sigma) | ||
| func5_gauss_i = zfit.pdf.Gauss(obs=func5_obs_i, mu=func5_mu_i, sigma=func5_sigma_i) | ||
| func5_prod_gauss_lst.append(func5_gauss_i) | ||
| func5_prod_gauss = zfit.pdf.ProductPDF(pdfs=func5_prod_gauss_lst[::-1]) |
There was a problem hiding this comment.
Move all of this inside a function
| compilable: bool = True, | ||
| verbose: bool = False, | ||
| compilable: bool = False, | ||
| verbose: int | str = 3, |
There was a problem hiding this comment.
| verbose: int | str = 3, | |
| verbose: int | str = 0, |
Let the default be quiet
Fixes #
Proposed Changes
Tests added
Checklist