TopK not supported?

[lucasjinreal]

onnxruntime.capi.onnxruntime_pybind11_state.NotImplemented: [ONNXRuntimeError] : 9 : NOT_IMPLEMENTED : Could not find an implementation for the node TopK(11)

onnxruntime                   1.1.0                  

Doesn't TopK supported for now?

[jywu-mysoft]

it may be due to the same issue #2814
which was recently fixed in master with
#2833
can you confirm if your model has topK which requires int64 type support?

[pranavsharma]

There's a followup PR that makes it easy to add more types here - #2842

[lucasjinreal]

@jywu-msft @pranavsharma thanks for your response. From onnx documentation:

Type Constraints
T : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double)

https://github.com/onnx/onnx/blob/master/docs/Operators.md#TopK

It should support input with type int64.

Actually, what surprise me is that I remmber I successfully forwarded such as likely model with TopK node. But now it fails, so strange.

[pranavsharma]

Which opset are you using? Can you share the model? I just updated topk to support int64_t for opset 11 only.

[lucasjinreal]

@pranavsharma I am using opset 11.

the model is here https://drive.google.com/open?id=1CegIKH7PF_HzGpzXuTkQ4r38cMyKIG5H

[cjermain]

I am also seeing this in opset 11 with onnxruntime 1.1.1, when trying to sort a Tensor of int64.

[pranavsharma]

@pranavsharma I am using opset 11.

the model is here https://drive.google.com/open?id=1CegIKH7PF_HzGpzXuTkQ4r38cMyKIG5H

This is an invalid model.

PS C:\Users\prs\work_projects\onnxruntime> .\build\Windows\Debug\Debug\onnx_test_runner.exe -vv C:\Users\prs\Downloads\centermask
2020-02-24 12:42:14.5223945 [E:onnxruntime:Default, runner.cc:185 RunTests] Test centermask failed:This is an invalid model. Type Error: Type 'tensor(bool)' of input parameter (1203) of operator (ReduceSum) in node () is invalid.
result:
        Models: 1
        Total test cases: 0
                Succeeded: 0
                Not implemented: 0
                Failed: 0
        Stats by Operator type:
                Not implemented(0):
                Failed:
Failed Test Cases:

[cjermain]

The following model reproduces the error with onnxruntime 1.1.2.

https://drive.google.com/file/d/1Vgj_evFx2mvUBIf3713OpKWnifpX7HlV/view?usp=drivesdk
(md5sum is c60f5519f3a6a8e0fc5a7c98641179c1)

I produced it with the below PyTorch 1.4.0 code, exporting to opset 11. I ran onnx.checker.check_model and did no see any errors (using onnx 1.6.0).

import torch
import torch.nn as nn

class SortNet(nn.Module):
    
    def forward(self, X):
        values, indices = X.sort(descending=True)
        return indices

net = SortNet()
X = torch.arange(10)

torch.onnx.export(
    net,
    X,
    'sort.onnx',
    export_params=True,
    do_constant_folding=True,
    opset_version=11,
    input_names=['input'],
    output_names=['output'],
    dynamic_axes={
        'input': {0: 'batch'},
        'output': {0: 'batch'},
    },
    verbose=True
)

[pranavsharma]

See #3092 (comment).