-
Notifications
You must be signed in to change notification settings - Fork 83
Closed
Labels
contribution welcomeWe welcome code contributions for thisWe welcome code contributions for thisgood first issueGood for newcomersGood for newcomersmodule: rewriter
Description
I've been using ONNXScript extensively lately and really appreciate how powerful and flexible it is. As part of that experience, I and @Johansmm would like to suggest a few optimization patterns that could improve performance and simplify common model graphs. These are generally applicable across many models:
- Fold
BatchNormalization
into preceding nodes (Conv
,ConvTranspose
,Gemm
) Rewriter: Fold Batchnorm nodes #2312 - Fuse
MatMul + Add
intoGemm
[Rewriter]: Add ∘ MatMul -> Gemm #2356 - Fuse
Min + Max
intoClip
[Rewriter]: add fusion rules for successive Min/Max patterns #2500 - Eliminate redundant
Reshape
/Flatten
nodes [rewriter] Unify reshape flatten ops #2518 - Fold
Pad
intoConv
/ConvInteger
[Rewriter] Add optimizer to fold Pad operators into Conv #2363 - Fuse successive
Clip
/ReLU
nodes [Rewriter]: fuse successive Relu/Clip nodes #2410
We're happy to contribute if any of these are a good fit. Feel free to add other suggestions or let us know if some of these are out of scope.
Metadata
Metadata
Assignees
Labels
contribution welcomeWe welcome code contributions for thisWe welcome code contributions for thisgood first issueGood for newcomersGood for newcomersmodule: rewriter