aadhithya / onnx-typecast

Script to typecast ONNX model parameters from INT64 to INT32.

onnx-typecast

A simple python script to typecast ONNX model parameters from INT64 to INT32.

Why?

I wanted to play around with ONNX.js and soon figured out that it doesn't support onnx models with INT64 parameters. Also, OpenCV doesn't seem to support INT64 parameters.

What does this script do?

• The script goes through the parameters of each node of the onnx model and blindly converts INT64 parameters to INT32.
• It also converts the constant parameters from INT64 to INT32.
• Finally, creates a new model and saves it.

What's the catch?

• The script does not handle overflows and blindly converts all INT64 parameters to INT32.
• So ops that require >INT32.max or <INT32.min values may not perform as expected.
• Please feel free to modify the script to account for this.

Alright. How do I use it?

• simple.
• Install requirements: pip install -r requirements.txt
• run: python convert.py path/to/int64_model.onnx path/to/converted_int32_model.onnx

Also Checkout

• microsoft/onnxjs#168 (comment)

onnx-typecast's Issues

Bad conversion of superpoint.onnx model

Hello,

I have tried to use your script (thanks for sharing!) to convert the model of superpoint.onnx (available here). Apparently the model is converted without any issue, but then when I try to load it with onnxruntime-web I get the following error:

Error: Can't create a session. ERROR_CODE: 1, ERROR_MESSAGE: Type Error: Type parameter (T) of Optype (Mul) bound to different types (tensor(int64) and tensor(int32) in node (/Mul_1).
    at t.checkLastError (bundle.min.js:1:307223)

On the other hand, if I use https://netron.app/ to check the model I can see that the "keypoints" output still has Int64 type :

I'm working with the following versions since I was not able to install the versions specified in the requirements:

numpy-1.26.1
typer-0.9.0
onnx-1.14.1
colorlog-4.7.2

Cast Ops in onnx file

Sometimes a onnx file has Cast Ops that cast data type to INT64. The case needs to be handled as well, by just modifying the attribute of Cast Ops to INT32.

Add CI and pypi release

• Add CI for listing, formatting and security.
• SemVer and release to PyPI.

This is an invalid model. Type Error: Type 'tensor(int32)' of input parameter is invalid.

Thanks for great work!

I'm getting this error after applying the conversion to SwinB transformer (when trying to use it for inference)

ort_sess = ort.InferenceSession('latest.i32.onnx')

  File "/python3.10/site-packages/onnxruntime/capi/onnxruntime_inference_collection.py", line 360, in __init__
    self._create_inference_session(providers, provider_options, disabled_optimizers)
  File "python3.10/site-packages/onnxruntime/capi/onnxruntime_inference_collection.py", line 397, in _create_inference_session
    sess = C.InferenceSession(session_options, self._model_path, True, self._read_config_from_model)
onnxruntime.capi.onnxruntime_pybind11_state.InvalidGraph: [ONNXRuntimeError] : 10 : INVALID_GRAPH : Load model from latest.i32.onnx failed:This is an invalid model. Type Error: Type 'tensor(int32)' of input parameter (7369) of operator (Reshape) in node (/model/backbone/patch_embed/Reshape) is invalid.

any idea what might be going wrong? or how to work around it ? :)

Conversion of MODNet .onnx model

Thanks for sharing this script, i tried it on MODNet. (background removal open source codebase)

I had an error running it, it worked but when trying to do inference i get an error due to the ReShape operator

INVALID_GRAPH : Load model from modnetpf16simint32.onnx failed:This is an invalid model. Type Error: Type 'tensor(int32)' of input parameter (480) of operator (Reshape) in node (Reshape_177) is invalid.

-