Translation of inputs to Resize, when using 'sizes' argument, drops required inputs

[kubaraczkowski]

Hi,

I found a bug happening when translating a Resize op that uses a 'sizes' argument.
The required input arguments 'rois' and 'scales' are being dropped, since they are None, but they are "required empty" by onnxruntime.

The strange thing is that the translation when running 'from function' is OK, but exporting the model misses the inputs.

Here is the minimal working example:

import numpy as np
import onnx
import onnxruntime as ort
from onnxscript import FLOAT
from onnxscript import opset20 as op
from onnxscript import script


@script()
def resize(X: FLOAT[800, 600]) -> FLOAT[512, 512]:
    return op.Resize(X, sizes=[512, 512])


# checker is OK
onnx.checker.check_model(resize.to_model_proto())

# This works - skipped arguments `roi` and `scales` get correctly translated to "empty tensors"
print("From function - OK")
X = np.eye(800, 600, dtype=np.float32)
Y = resize(X)

# This doesn't - the `roi` and `scales` get omitted, which makes `sizes` a second argument, while it needs to be fourth.
try:
    print("\n\n\nBug - not OK \n==================")
    session = ort.InferenceSession(resize.to_model_proto().SerializeToString())
    Y = session.run(None, {"X": X})
except Exception as e:
    print(e)
    print("==================")

# workaround - add two empty arguments for `roi` and `scales`
args = resize.function_ir.stmts[-1].args
resize.function_ir.stmts[-1].args = (args[0], "", "", args[1])

# Now it's OK
print("\nWith workaround - OK")
session = ort.InferenceSession(resize.to_model_proto().SerializeToString())
Y = session.run(None, {"X": X})

And the output of it:

From function - OK

Bug - not OK 
==================
[ONNXRuntimeError] : 10 : INVALID_GRAPH : This is an invalid model. Type Error: Type 'tensor(int64)' of input parameter (const) of operator (Resize) in node (n1) is invalid.
==================

With workaround - OK

I'm using onnxscript=0.1.0.dev20250108

And lastly - THANKS - onnxscript is such a great idea for writing pre/post-processing functions!

[justinchuby]

Thanks for reporting. We will look into this