Justin Chu comments

Results 452 comments of


                                            Justin Chu

WIP: Release pipeline without offline steps (PR to discuss the ideas)

The diagram flow looks very good and clean, thanks!

Want to substitute Expand operator with some other operator in ONNX due to compatibility issues with hardware

You may use the rewriter from onnxscript for this purpose: https://onnxscript.ai/tutorial/rewriter/index.html

Batching in SequenceLength Operator

Do you have variable length sequences? Otherwise batching can happen within the sequence?

Split onnx model in architecture and weights

Yes. Simply call onnx.save with external data = True will do.

Shape inference does not produce shapes for every node

ONNX 1.9 is outdated. Please use the latest version by installing onnx-weekly and let us know

Shape inference does not produce shapes for every node

Please remove onnx, onnx-weekly and reinstall onnx-weekly. I suspect there’s an installation conflict because onnx.load should always work

Convert a model with custom pytorch CUDA kernel

Maybe this? https://onnxruntime.ai/docs/reference/operators/add-custom-op.html

No Adapter From Version $14 for Mul

Backward conversion is not complete. You may register it here https://github.com/onnx/onnx/blob/99e1fd352c05c3176770080824fd7a8c474c97c0/onnx/version_converter/convert.h#L511 and send a PR. Thanks.

Improve test coverage by convering model before calling onnxruntime (2)

We previously removed this behavior because it tends to be indirect and opaque. Trying to clarify for my understanding - I wonder what the goal is for these tests? If...

Result accuracy is different with PyTorch

We consider this reasonable within the tolerance. It is not possible to produce exactly matching floating point results due to how they are computed. Did you observe any difference in...