op.Resize parameter sizes cannot accept Type tensor(int64)

[vincentme]

behavior

The sizes parameter computed on the fly works in onnxscript eager-mode. But the converted model run by onnxruntime complains

InvalidGraph: [ONNXRuntimeError] : 10 : INVALID_GRAPH : This is an invalid model. Type Error: Type 'tensor(int64)' of input parameter (new_shape) of operator (Resize) in node (n4) is invalid.

However, according to the doc of onnx, input sizes should indeed be a tensor(int64).

code to reproduce

import numpy as np
from onnxscript import script, FLOAT, opset20 as op


@script()
def test(data: FLOAT[None, None]) -> FLOAT[None, None]:
    shape = op.Shape(data)
    new_shape = shape/2
    print(new_shape)
    return op.Resize(data, sizes=new_shape)

data = np.random.randn(16, 32).astype(np.float32)
res_onnxscript = test(data)
model = test.to_model_proto()

from onnxruntime import InferenceSession
sess = InferenceSession(model.SerializeToString())
feeds = {'data': data,}
outpus_onnxruntime = sess.run(None, feeds)

[justinchuby]

I speculate it is because onnx script is not supplying the arguments correctly. For now, I would supply inputs only positionally and supply attributes as keyword arguments. Omitted inputs can be represented as None.

[vincentme]

I speculate it is because onnx script is not supplying the arguments correctly. For now, I would supply inputs only positionally and supply attributes as keyword arguments. Omitted inputs can be represented as None.

I see. Fixed with op.Resize(data, None, None, new_shape)

Thanks!

[ebarkhordar]

Justin Chu (@justinchuby) your guess holds up. At main 08898ad0 onnxscript builds the input list by appending, so an omitted optional input is never materialised and everything behind it slides down a slot. _internal/param_manipulation.py:73 appends kwargs[param.name] at the current end instead of at the parameter's index, so Resize(data, sizes=...) emits ['data', 'new_shape'] and new_shape lands in roi.

Resize is the lucky case, because roi is float and ORT throws. Where the slots share a type nothing complains:

op.Clip(x, max=Constant(2.0))  ->  Clip node inputs ['x', 'tmp']
x = [-5 -1 0 1 5]              ->  [2 2 2 2 5]      that is min=2, not max=2

Valid model, clean run, wrong numbers. 19 default-domain ops declare an optional input with a later input behind it, so it is not only Resize.

Reads like the same bug as #2007, and the0cp's #2961 fixes it. I ran both cases on that branch: Resize comes out ['data', '', '', 'new_shape'] and runs, Clip comes out ['x', '', 'tmp'] and clips above, and Clip(x, min=...) is untouched.