After adding preprocessing steps to the model, on inference an error throws: Non-zero status code returned while running Add node.

[feff2]

Describe the issue

I added preprocessing steps for TrOCR but now I get the following error: onnxruntime.capi.onnxruntime_pybind11_state.RuntimeException: [ONNXRuntimeError] : 6 : RUNTIME_EXCEPTION : Non-zero status code returned while running Add node. Name:'/embeddings/Add' Status Message: C:\a_work\1\s\onnxruntime\core/providers/cpu/math/element_wise_ops.h:560 onnxruntime::BroadcastIterator::Append axis == 1 || axis == largest was false. Attempting to broadcast an axis by a dimension other than 1. 386 by 578

To reproduce

This is my code, how i add it:
model = onnx.load("trocr_onnx/encoder_model.onnx")

inputs = [create_named_value('image', onnx.TensorProto.UINT8, ['num_bytes'])]
pipeline = PrePostProcessor(inputs, 18)

pipeline.add_pre_processing(
[
ConvertImageToBGR(),
Resize((384, 384), policy='not_larger'),
ChannelsLastToChannelsFirst(),
ImageBytesToFloat(),
Normalize([(0.5, 0.5), (0.5, 0.5), (0.5, 0.5)]),
Unsqueeze([0]),
]
)

modified_model = pipeline.run(model)
onnx.checker.check_model(modified_model)
onnx.save_model(modified_model, 'trocr_encoder.onnx')

Urgency

No response

Platform

Windows

OS Version

10

ONNX Runtime Installation

Built from Source

ONNX Runtime Version or Commit ID

1.18.0

ONNX Runtime API

Python

Architecture

X64

Execution Provider

Default CPU

Execution Provider Library Version

No response

[skottmckay]

Hard to say without seeing the model. What was the expected input shape of the original model?

[feff2]

expected input shape is 1x3x384x384

[skottmckay]

Most likely your issue is this: Resize((384, 384), policy='not_larger'),

If the input image is square it should work. If it is not square, after the Resize the longest size will be 384 but the smallest will be less than that.

Typically you would do a centered crop or letterbox the resized image so that the original aspect ratio is maintained and the input is 384x384 as the model requires.

[feff2]

thanks!