Make convolve mode symbolic to avoid unnecessary large convolution in gradient graph #1522
+279
−276
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
ricardoV94
force-pushed
the
make_convolve_mode_symbolic
branch
3 times, most recently
from
48a5ce3 to
95b4cb3
Compare
ricardoV94
force-pushed
the
make_convolve_mode_symbolic
branch
3 times, most recently
from
281bbf9 to
fe2ea6b
Compare
ricardoV94
force-pushed
the
make_convolve_mode_symbolic
branch
2 times, most recently
from
12e3123 to
5be6968
Compare
Codecov ReportAttention: Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## main #1522 +/- ##
==========================================
- Coverage 82.04% 82.04% -0.01%
==========================================
Files 231 230 -1
Lines 52364 52346 -18
Branches 9217 9212 -5
==========================================
- Hits 42962 42947 -15
- Misses 7094 7095 +1
+ Partials 2308 2304 -4
🚀 New features to boost your workflow:
|
jessegrabowski
approved these changes
Jul 8, 2025
| in2_bar = in2_bar[nmk : in2_bar.shape[0] - nmk] | ||
|
|
||
| return [in1_bar, in2_bar] | ||
| # If mode is "full", or mode is "valid" and k >= n, then in1_bar mode should use "valid" convolve |
| out[0] = s.flatten()[0] | ||
| def perform(self, node, inputs, output_storage): | ||
| # not using .item() because that returns a Python scalar, not a numpy scalar | ||
| output_storage[0][0] = inputs[0][()] |
There was a problem hiding this comment.
That [()] syntax is really ugly
There was a problem hiding this comment.
It's how you index/update a 0d array in numpy
There was a problem hiding this comment.
You can also use None I believe.
Edit: It only works for assignment. If this is really what they want you to do in this case, that sucks. But c'est la vie.
There was a problem hiding this comment.
The idea is you're indexing an array without indices / dimensions, so you can't say pick first entry of first dimension (there are no dims). Anyway surprised it bothers you too much
ricardoV94
force-pushed
the
make_convolve_mode_symbolic
branch
from
5be6968 to
9026dd8
Compare
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Instead of statically parametrizing the convolution type at the Op level, it now uses a scalar boolean that can be set symbolically. This was fine except for JAX, where we can't branch symbolically like that.
As long as the mode is constant (which is the case when the original convolve was full, or the core shapes are statically known) this should be fine, except for one hiccup. In the dispatch of Blockwise we couldn't see the outer inputs.
I added some functionality when we create the dummy core node to propagate inputs if these don't have batch dimensions, which means they won't change over iterations and so making compile or rewrite decisions based on this should be safe. This can also be used for infer_shape for instance, which could help with lowering some Ops to numba
These changes would also allow us to compile a constant convolve mode in C/Numba, but benchmarks didn't show any gains so I didn't bother doing that. In any case, future implementations of Blockwise for certain Ops can make use of that information.
Relevant benchmark tests:
Note the worst case scenario when we were doing a full convolution for the smaller input in the gradient of a valid convolution.
📚 Documentation preview 📚: https://pytensor--1522.org.readthedocs.build/en/1522/