Skip to content

Weird behaviour of Gaussian Linear Task in mini SBI-BM #1781

@jsvetter

Description

@jsvetter

🐛 Bug Description

The Gaussian Linear task in the mini sbibm seems to be bugged.

🔄 Steps to Reproduce

Python version: Python 3.11.2
SBI version: 0.25.0

To reproduce, run
pytest tests/bm_test.py --bm --bm-num-simulations XXX
at various simulations budgets.

E.g. below, I did it for 100, 2000 (default), and 10000 with NPE-PFN and NPE using NSF.

Both methods are very bad, and standard NPE even gets worse with more simulations.

NPE with NSF:
100 simulations: 0.787
2000 simulations: 0.889
10000 simulations: 0.927

NPE-PFN is usually super good at this task, but is at around 0.95 at all three budgets.

✅ Expected Behavior

With more simulations performance should improve. And NPE-PFN should be very good at this in terms of C2ST with even very few simulations, and not around 0.95.

Metadata

Metadata

Assignees

No one assigned

    Labels

    bugSomething isn't working

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions