pymc-devs / pymc-devs/pytensor
Rewrite `expand_dims` implied in vector indices
Open
Nobody has claimed this yet.
graph rewriting
indexing
numba
- Dominant language
- Python
- Stars
- 644
- Forks
- 208
- Avg merge
- 2d 14h
- Merged PRs (30d)
- 16
Description
Description
The model behind #1132 cannot run in non-obj numba due to an implicit expand_dims in the vector indexing, that looks like:
import pytensor
import pytensor.tensor as pt
import numpy as np
a = pt.tensor(shape=(None, None))
b = a[pt.arange(10)[:, None], pt.arange(10)[:, None]]
c = a[pt.arange(10), pt.arange(10)][:, None]
# Issues UserWarning: Numba will use object mode to run AdvancedSubtensor's perform method
fn_b = pytensor.function([a], b, mode="NUMBA")
# Runs in non-obj mode
fn_c = pytensor.function([a], c, mode="NUMBA")
test_a = pt.random.normal(size=(10, 10)).eval()
np.testing.assert_allclose(fn_b(test_a), fn_c(test_a))
Contributor guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
Start with the vector-indexing reproducer in the issue and the AdvancedSubtensor path it exercises. Compare the implicit expand_dims case with the explicit trailing-axis case, then run the shown NUMBA compilation and equivalence checks. Done means the vector-indexing form compiles in non-object Numba mode and matches the explicit form.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python
- Domain
- compilers
- Issue type
- Refactor
- Difficulty
- 4/5
- Estimated time
- 3-5 days
- Activity status
- Stale
- Clarity
- Mostly clear
- Newbie friendliness
- 35/100