"AliasWithName is not a registered function/op" when run converted onnx model

[hieubz]

I used tools/deploy/caffe2_converter.py to convert to onnx model. But when I initiated this model by onnxruntime on google colab, this throwed an exception:

[ONNXRuntimeError] : 1 : FAIL : Fatal error: AliasWithName is not a registered function/op

My code to load onnx model

`
def to_numpy(tensor):
return tensor.detach().cpu().numpy() if tensor.requires_grad else tensor.cpu().numpy()

class ONNX_DETECT:

def __init__(self, onnx_model_path, device):
    self.onnx_model_path = onnx_model_path
    self.device = device.upper()
    print(' ---- running on ', onnxruntime.get_device())

    self.detector = onnxruntime.InferenceSession(self.onnx_model_path)

def detect(self, input_image):

    inputs = {self.detector.get_inputs()[0].name: to_numpy(input_image)}
    outputs = self.detector.run(None, inputs)

    return outputs

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
detectron_model = ONNX_DETECT('/content/model.onnx', device.type)
`
Have anyone faced this issue?

[ppwwyyxx]

As https://detectron2.readthedocs.io/modules/export.html#detectron2.export.Caffe2Tracer.export_onnx says

ote that the exported model contains custom ops only available in caffe2, therefore it cannot be directly executed by other runtime. Post-processing or transformation passes may be applied on the model to accommodate different runtimes.

so this is working as expected

[hieubz]

Would you please give me some ideas to deal with it? I still vague about that, especially how to run detectron2 on onnxruntime.
Thanks.

[hieubz]

I saw that there are 3 options when I export model (caffe2, onnx, torchscript), why exported model only available in caffe2 ? why onnx exported model can not run on onnxruntime but needs post-processing ?

[menchunlei]

I saw that there are 3 options when I export model (caffe2, onnx, torchscript), why exported model only available in caffe2 ? why onnx exported model can not run on onnxruntime but needs post-processing ?
@PhamDuyHieutb
because op isn't implemented in the backend of onnxruntime.

[Muhammad-Talha-MT]

I am facing the same issue
What is Post-processing or transformation passes?