Help Wanted: Implementing ONNX Ops

[antimora]

The Burn community would greatly appreciate your help in completing the missing ONNX operations in the burn-import crate. By contributing to this effort, you can help expand the functionality and usability of Burn for a wider range of machine learning models.

List of Missing Ops (Available in Burn but not in burn-import)

Top Requested Ops

• Expand
• Tile (dependent on #1715)
• Slice

Relatively Easy Ops (Similar to Existing Implemented Ops)

• And
• ArgMin
• ArgMax
• Attention
• AveragePool1d
• BitShift
• BitwiseAnd
• BitwiseNot
• BitwiseOr
• BitwiseXor
• Ceil
• ConvTranspose1d
• EyeLike
• Floort
• Gather
• Greater
• GreaterOrEqual
• Less
• LessOrEqual
• Max
• MaxPool1d
• Mean
• Min
• NonZero
• Not
• Or
• PRelu
• Pad
• Range
• ReduceMin
• ReduceProd
• ReduceSum
• Round
• Scatter
• Size
• Squeeze
• Sum
• TopK
• Trilu
• XOr
• IsNaN
• Bernoulli

Harder Ops (Not Similar to Existing Implemented Ops)

• GRU
• GroupNormalization
• If (dependent on #724)
• InstanceNormalization
• LSTM
• MatMulInteger
• OneHot
• RNN
• RandomNormal
• RandomNormalLike
• RandomUniform
• RandomUniformLike
• Resize

For a comprehensive list of all supported ONNX Ops, please refer to the SUPPORTED-ONNX-OPS.md file in the Burn repository.

Getting Started

To begin contributing, please follow the instructions in the contributor book:

# Build the book and open in a browser from the project root
cargo xtask books contributor open

Once the book is built, navigate to http://localhost:3011/guides/onnx-to-burn-conversion-tool.html for detailed guidance on how to implement missing ONNX operations in the burn-import crate.

Related Issues Submitted by Users

Several users have submitted issues related to missing ONNX operations. These issues can provide valuable context and help prioritize the implementation of specific ops:

1. Issue #1593
2. Issue #1560
3. Issue #1544
4. Issue #675
5. Issue #673
6. Issue #600
7. Issue #510

By tackling these missing ops, you can help address the needs of the Burn community and contribute to the growth and effectiveness of the project. Thank you for your interest in contributing to Burn!

[Luni-4]

@antimora

Can we add the Split operator to the list too? Looking at this list, it's not present in burn as well though 🤔

[antimora]

@Luni-4

@antimora

Can we add the Split operator to the list too? Looking at this list, it's not present in burn as well though 🤔

This ticket is mostly catching up with ONNX ops that can be implemented immediately without adding a Burn OP.

I think we should file a separate ticket (if you don't mind filing since you might know a better use case)

[Luni-4]

@antimora

You are perfectly right, I should have noticed the difference. Yes, I can file another issue

[steckes]

Hey I am just waiting for GRU ONNX import, any plan on implementing this?
How much work is the import, as the implementation should be done already according to the Ops page?

[laggui]

I don't think anyone is actively working on adding GRU to onnx import.

I'm not familiar with the differences with the GRU ONNX spec, I'd have to look into it to see how easy it is for someone to add support for this op.

[sebimarkgraf]

I am having issues when important torch onnx exports to burn.

Could it be that the ONNX import is missing support for ONNX functions?
https://onnx.ai/onnx/intro/concepts.html#functions

[antimora]

I am having issues when important torch onnx exports to burn.

Could it be that the ONNX import is missing support for ONNX functions? https://onnx.ai/onnx/intro/concepts.html#functions

If you're trying to import ONNX file with functions, then unfortunately we do not support it yet. We have an outstanding ticket for this: #723.

[NewBornRustacean]

Good morning/evening! Thanks for this awesome framework and kind explanation/documentation 👍
I'm trying to add MatMulInteger( draft pr is here: #2846 ), so far code generation parts are almost done.

will add some other functions/tests and turn the PR into ready to review!

[antimora]

@NewBornRustacean

Good morning/evening! Thanks for this awesome framework and kind explanation/documentation 👍 I'm trying to add MatMulInteger( draft pr is here: #2846 ), so far code generation parts are almost done.

will add some other functions/tests and turn the PR into ready to review!

Thank you for your contribution! I will take a look today!

[AshAnand34]

I am currently working on the bitwise operators for this issue. I followed the steps in the contributor book to create all of them and add them to burn-import. While attempting to create the test ONNX model for the bitwise not operator inside onnx-tests, I ran into the following error:

torch.onnx.errors.SymbolicValueError: ONNX export does NOT support exporting bitwise Not for non-boolean input values  [Caused by the value '0 defined in (%0 : Int(4, strides=[1], requires_grad=0, device=cpu) = prim::Param()
)' (type 'Tensor') in the TorchScript graph. The containing node has kind 'prim::Param'.]

    Inputs:
        Empty
    Outputs:
        #0: 0 defined in (%0 : Int(4, strides=[1], requires_grad=0, device=cpu) = prim::Param()
    )  (type 'Tensor')

[laggui]

If the node cannot be exported via pytorch, the graph can be manually constructed using onnx utils. See for example this gather test.

[AshAnand34]

If the node cannot be exported via pytorch, the graph can be manually constructed using onnx utils. See for example this gather test.

The PyTorch export approach worked for the bit shift and the bitwise and operators, but not for bitwise not, which is the weird part (changed opset_version to 18 as it is the only version of these operators in ONNX).

[antimora]

If the node cannot be exported via pytorch, the graph can be manually constructed using onnx utils. See for example this gather test.

The PyTorch export approach worked for the bit shift and the bitwise and operators, but not for bitwise not, which is the weird part (changed opset_version to 18 as it is the only version of these operators in ONNX).

If OpSet 18 works, then go ahead use it.

[AshAnand34]

If the node cannot be exported via pytorch, the graph can be manually constructed using onnx utils. See for example this gather test.

The PyTorch export approach worked for the bit shift and the bitwise and operators, but not for bitwise not, which is the weird part (changed opset_version to 18 as it is the only version of these operators in ONNX).

If OpSet 18 works, then go ahead use it.

I tried manually constructing the graph via onnx based on what @laggui suggested and it worked. Thank you so much.

[AshAnand34]

I added some tests to the test_onnx.rs file for the bitwise operations. After running cargo clippy --all, I got the following errors that I am unsure of how to fix:

error[E0308]: mismatched types
--> C:\Users\aashi\OneDrive\Documents\MiscellaneousCoding\burn\target\debug\build\onnx-tests-4127cf55dd0a9cbe\out/model/bitshift_left.rs:50:41
|
50 | let pow1_out1 = cast1_out1.powf(constant1_out1);
| ---- ^^^^^^^^^^^^^^ expected Tensor<B, 1>, found f32
| |
| arguments to this method are incorrect
|
....
Can you please let me know what I need to do to resolve these issues?

[antimora]

@AshAnand34, please submit a draft PR and we can discuss over there.