separating onnx parsing from burn-import

[skewballfox]

Pull Request Template

Checklist

• confirm we want the name to be onnx-ir
• deduplicate and remove unused dependencies
• Confirmed that run-checks all script has been executed.
• Made sure the book is up to date with changes in this PR.

Related Issues/PRs

#1812

Changes

Separated out all the onnx specific code in burn-import to a new crate onnx-ir (which we can rename to whatever), and made some necessary changes to burn-import since OnnxGraph and other ir structs are now external. I also aliased a few types to just better distinguish between onnx and burn types.

some code had to be duplicated (flatten_config)or made public for the sake of using between both (constant_conversion)

Testing

so far I've tested with onnx-test and burn-import. All test seem to be passing. examples seem okay

I think we should rework how onnx test are being used so that the models, and test themselves could be potentially reused. I think we can do that separately from this PR though.

Python side that would mean serializing graph inputs and outputs, burn side I'm guessing some macros for op types (such as binary_op_test!) though I'm open to suggestions if you guys want to handle it another way.

[codecov]

Codecov Report

Attention: Patch coverage is 94.58333% with 13 lines in your changes missing coverage. Please review.

Project coverage is 84.86%. Comparing base (755c070) to head (b01c11e).

Files	Patch %	Lines
crates/onnx-ir/src/util.rs	68.75%	10 Missing ⚠️
crates/burn-import/src/onnx/to_burn.rs	98.21%	3 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1921      +/-   ##
==========================================
- Coverage   84.87%   84.86%   -0.01%     
==========================================
  Files         797      798       +1     
  Lines       95082    95105      +23     
==========================================
+ Hits        80697    80711      +14     
- Misses      14385    14394       +9     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[antimora]

Looks good. A minor comment to review/clean up the dependencies.

But I think it's almost ready to be approve.

Since this is a brand new crate, we also need to update publish.yml under .github.yml.

Also, can you please double check adding-a-new-operation-to-burn.md is up to date? Not sure if we have any references to about the onnx portion.

@nathanielsimard @laggui can you sign off on the name?

[laggui]

onnx-ir sounds good to me! Will try to take some time to review this.

[skewballfox]

Also, can you please double check adding-a-new-operation-to-burn.md is up to date? Not sure if we have any references to about the onnx portion.

I rewrote it, but I'm having trouble figuring out the best way to word it. I'd definitely appreciate a proof reader on the changed section.

Also I added a readme though it's currently pretty barebones still figuring out what information it should have for other implementors.

[laggui]

Totally forgot about this PR, I'll hold off on merging an new ONNX ops if we're going to merge this soon otherwise you'll always have to fix the conflicts 😅

Otherwise, the refactor LGTM.

[antimora]

Approving in advance. There are a couple minor text fixes but looks good overall.

[antimora]

Merged! Thanks for your work. I hope this crate will be useful to other project as well.