Bug: Error converting Resize operator with scales/sizes as inputs (opset 11+)

[lksrz]

Hi @Oil3,
Thanks for creating this converter!
While converting an ONNX model (yolo6's pt exported to onnx in my case, but likely applicable to others) using the convert.py script, I encountered an error during the processing of a Resize operator (opset 11 or later).
Error:
Initially, an UnboundLocalError: local variable 'scale' referenced before assignment occurred within the _convert_resize function in _operators_nd.py. This happened because the scales parameter for the Resize operation was provided as an input tensor (standard for opset 11+) rather than an attribute. The code attempted to access the scale variable before it was assigned a value from the input tensor.
During the fix attempt, an AttributeError related to graph.has_initializer also occurred, indicating the need to use node.input_tensors for accessing input values.
Root Cause:
The current implementation of _convert_resize doesn't fully handle the case where scales or sizes are provided as input tensors (indices 2 and 3 respectively in opset 11+) instead of attributes (opset 10 style).
Proposed Solution:
Modify _convert_resize to first check node.input_tensors for the presence of scales and sizes input names. If found, retrieve the values from there. If not found, then fall back to checking the node attributes (node.attrs.get('scales', None)).
I have tested a fix locally which resolves this issue for my model. I will link a Pull Request.
Best!