Got "NOT_IMPLEMENTED" error on OneHot with int32

[disktnk]

Describe the bug

When call OneHot with int32 indeices, get "NOT_IMPLEMENTED" RuntimeError.

System information

• OS Platform and Distribution: Ubuntu18.04
• ONNX Runtime installed from: binary via PyPI
• ONNX Runtime version: 0.4.0
• Python version: 3.7.2
• Visual Studio version (if applicable):
• GCC/Compiler version (if compiling from source):
• CUDA/cuDNN version:
• GPU model and memory:

To Reproduce

based https://github.com/onnx/onnx/blob/v1.5.0/onnx/backend/test/case/node/onehot.py#L47

    axisValue = 1
    on_value = 3
    off_value = 1
    output_type = np.float32
    node = onnx.helper.make_node(
        'OneHot',
        inputs=['indices', 'depth', 'values'],
        outputs=['y'],
        axis=axisValue,
    )
    indices = np.array([[1, 9],
                        [2, 4]], dtype=np.int32)  # original np.float32
    depth = np.array([10], dtype=np.float32)
    values = np.array([off_value, on_value], dtype=output_type)
    y = one_hot(indices, depth, axis=axisValue, dtype=output_type)
    y = y * (on_value - off_value) + off_value

        if isinstance(path_or_bytes, str):
            self._sess.load_model(path_or_bytes)
        elif isinstance(path_or_bytes, bytes):
>           self._sess.read_bytes(path_or_bytes)
E           RuntimeError: [ONNXRuntimeError] : 9 : NOT_IMPLEMENTED : Could not find an implementation for the node OneHot(9)

Expected behavior

No error, OneHot allows int32 input by Operators.md

Screenshots

Additional context

[disktnk]

Ah, I miss PR #1317 , so I wait next release, thx!

[disktnk]

I tested them with version 0.5.0, and got same error on OneHot op with int32 type.

Case1:

• indices: int64
• depth: int64

--> no error, works well, thanks!

Case2:

• indices: int32
• depth: int32

--> got "NOT_IMPLEMENTED" error, I think the condition is acceptable on OneHot op.

Case3:

• indices: float32
• depth: int64,int32

or

• indices: int64,int32
• depth: float32

--> also got "NOT_IMPLEMENTED" error, but I don't know runtimes had better to implement the condition or not. The docs says "In case 'indices' is of non-integer type, the values will be casted to int64 before use", so naturally the conditions are acceptable I think.

[hariharans29]

Hi @disktnk ,

Thanks for bringing this up.

Currently, the OneHot op doesn't support all type combinations.

The PR you are referring to (#1317) supported in_type: int64, out_type: float, depth: int64.

1. From what I understand from your code python snippet, you are looking for in_type: int32, out_type: float, depth: float. Is this correct ?

2. What other combinations would you like? At this time, I don't think we plan to support all type combinations that are possible as per spec. We would like the type support to be based on intuitive use-cases and request-driven.

Thanks

[disktnk]

Thanks for reply! It's reasonable and convinced your policy

We would like the type support to be based on intuitive use-cases and request-driven."

Answer:

1 -> Yes, and the snippet just for reproducible
2. -> We're looking for in_type: int32, depth: int32, out_type: float32

[hariharans29]

Thanks for reply! It's reasonable and convinced your policy

We would like the type support to be based on intuitive use-cases and request-driven."

Answer:

1 -> Yes, and the snippet just for reproducible
2. -> We're looking for in_type: int32, depth: int32, out_type: float32

Thanks for the confirmation. So there are 2 combinations requested -

1. in_type: int32, out_type: float, depth: float
2. in_type: int32, out_type: float, depth: int32

Will send a PR out.

[hariharans29]

Hi @disktnk - The requested type support should be available if you build from master. Thanks.

[disktnk]

Great!! I'll check it later, thx!