ReduceSum node not implemented

[tanguycdls]

Describe the bug
To create a test to check an inference wrapper around OnnxRuntime I created a really simple model adding the inputs and expanding dims as so:

class model(nn.Module):
    def __init__(self):
        super().__init__()
    def forward(self, x):
        return x.sum(dim=1).reshape(-1, 1).expand(-1, 10)

System information

• Linux Debian
• binary pip
• 0.5 and tested on 0.4
• 3.7 and tested 3.6
• Visual Studio version na
• GCC/Compiler version na
• CUDA/cuDNN version: na
• GPU model and memory: na

To Reproduce
Export the model from Pytorch using

from torch import onnx
a = model()
onnx.export(a, torch.from_numpy(np.array([1, 2, 3, 4])), 'model.onnx')
b = InferenceSession('model.onnx')

Expected behavior
We should be able to load since ReduceSum is implemented in the list of operators.
'''
RuntimeError: [ONNXRuntimeError] : 9 : NOT_IMPLEMENTED : Could not find an implementation for the node ReduceSum(1)
'''

[faxu]

See this page for the list of supported ops and types: https://github.com/microsoft/onnxruntime/blob/master/docs/OperatorKernels.md

What is the datatype for the ReduceSum op in your model? Current support types in ORT are:
T = tensor(int32), tensor(float), tensor(double)

[tanguycdls]

Hi thanks for the answer !
it's indeed int64 type (Long in Pytorch) I checked and its compatible with ONNX itself (int64). Maybe it would worth having a more verbose explanation when getting ReduceSum not implemented and adds the dtype for which it's not implemented ?

Thanks for the help,