Unsupported left operand type for - ("FLOAT") [operator]

[tuanle0910]

Hi everyone, I am trying to use onnxscript to convert a function into a script via the @script decorator. I have two questions:

1. Is there any ways to do subtraction between an input with the FLOAT type and the input with np.float32? If not, how do I convert from one to the other (I have looked up ALL the possible documentations from onnx but it's not much helpful). Right now I keep getting the error error: Unsupported left operand type for - ("FLOAT") [operator] with note: Right operand is of type "BFLOAT16 | BOOL | COMPLEX128 | COMPLEX64 | DOUBLE | <11 more items>"
2. Can I still use the normal operator (e.g. +, -, :, *) to do addition/subtraction/division/multiplication within the function that is supposed to be rewritten into .onnx model? The reason I ask is because it looks like onnxscript needs data to be of TensorType?

Package versions. onnxscript = 0.1.0.dev2024112, onnx=1.17.0, onnxconverter-common=1.13.0, onnxruntime=1.20.0, skl2onnx=1.17.0

Example.

         import numpy as np
         import onnx
         import pandas as pd
         from onnxscript import FLOAT
         from onnxscript import opset15 as op
        from onnxscript import script      
        z1 = 5.0
        z2 = 3.0
        t1 = z1.astype('np.float32')
        t2 = z2.astype('np.float32')
        cum_mins = [1,2,3,4]
        @script()
        def model_1(x: FLOAT, y: FLOAT) -> Tuple[Any, Any]:
            # Define constants
            coeff1 = op.Constant(value_float=t1)
            coeff2 = op.Constant(value_float=t2)

            mean_x = op.Constant(value_float=x)
            mean_y = op.Constant(value_float=y)

            exp_arg = coeff1 * (x - mean_x) + coeff2 * (y - mean_y)    
            exp = op.Constant(value_float=0.75) * op.Exp(exp_arg)

            h = op.Constant(value_floats=cum_mins)

            return (exp, h)

[tuanle0910]

@justinchuby I'm sorry, but can someone please help me with my question above? I kept encountering these errors while running make lints with mypy ---> Not sure if this is due to the incompatibility between onnxscript and mypy? Without running lints check, the code works perfectly.

For example, make lints complained the following error: Argument 1 to "Where" of "Opset16" has incompatible type "bool"; expected "BOOL" [arg-type], but I tried it in the Python Console within Pycharm and it worked perfectly well --- so I don't understand why it complains this, and how to circumvent it. Can you please help check on your end the compatibility between mypy and onnxscript?

x = 2.0
op.Where(x==1.0, 0.0, 3.0)

[titaiwangms]

1. Use op.Cast to cast the value to your intended datatype first
2. Yes

[tuanle0910]

@titaiwangms Thank you very much for your help, Mr. Wang. I tried op.Cast but I still got the error message of Left Operand Type is not supported for '*' --- Is there an internal issue between the data type used by onnxscript and the one by mypy?

[justinchuby]

The current typing system we have in onnxscript doesn't work well with mypy. I suggest disabling mypy for files that uses onnxscript for now. That's something on our roadmap but may take time.