feat: resize onnx import

[mosure]

checklist

• 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

• onnx op import sprint: Help Wanted: Implementing ONNX Ops #1714

changes

support Resize onnx import

testing

cargo test -p burn-import --color=always -- --color=always
cargo test -p onnx-tests --color=always

[mosure]

I'm looking for feedback to solve:

onnx resize sizes tensor -> burn interpolate size constant

when parsing the resize node, sizes input is a tensor, however, I have not found a method to do constant folding over the sizes input tensor, burn interpolate requires size constants:

pub fn interpolate<R: JitRuntime, E: JitElement + Element>(
    input: JitTensor<R, E, 4>,
    output_size: [usize; 2],
    options: InterpolateOptions,
) -> JitTensor<R, E, 4> {
...

[laggui]

In the codegen for the node you could get the values in the size tensor. Something like

#size.to_data().value

/edit: shoot I didn't realize the output_size is a fixed size array, that won't work.

[mosure]

In the codegen for the node you could get the values in the size tensor. Something like

#size.to_data().value

/edit: shoot I didn't realize the output_size is a fixed size array, that won't work.

it is a good idea, something like:

let output_size: [usize; 2] = #output_size.to_data()
    .value
    .iter()
    .map(|x| x.elem::<i64>() as usize)
    .collect::<Vec<usize>>()
    .try_into()
    .unwrap();

[laggui]

Yeah something like that could work to preserve the flexibility.

I think in most cases the sizes could be constant so we might be able to get the values directly when parsing the graph, but not entirely sure.

I see in the spec that ONNX also supports providing a scales instead, but let's limit the scope of this PR. We can just parse the other unsupported fields/inputs and throw an error message for the cases we don't support.

[codecov]

Codecov Report

Attention: Patch coverage is 95.55556% with 10 lines in your changes missing coverage. Please review.

Project coverage is 86.14%. Comparing base (671ec8c) to head (6e00ba9).

Files	Patch %	Lines
crates/burn-import/src/onnx/dim_inference.rs	80.95%	4 Missing ⚠️
crates/burn-import/src/onnx/op_configuration.rs	75.00%	4 Missing ⚠️
crates/burn-import/src/burn/node/resize.rs	98.67%	2 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1863      +/-   ##
==========================================
+ Coverage   86.11%   86.14%   +0.03%     
==========================================
  Files         777      778       +1     
  Lines       90555    90780     +225     
==========================================
+ Hits        77979    78202     +223     
- Misses      12576    12578       +2     

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

[Luni-4]

Thanks a lot @mosure! It's fine for me!

I'm going to wait for a final review by @laggui though, just to be sure I haven't forgotten something

[laggui]

LGTM! Just a minor request to remove the commented dead code

I think this is probably the operator with the most attributes and optional inputs I've seen thus far 😅

Glad you didn't try to cover every use case because we can't support them anyway hahah

/edit: oh and we need to fix the merge conflicts