No importer registered for op: Einsum

AI & Data Science Deep Learning (Training & Inference) TensorRT
1

Description

I execute this command
TensorRT-7.0.0.11/bin/trtexec --onnx=action.onnx --saveEngine=action.trt

But failed

Environment

TensorRT Version: 7.0.0.11
GPU Type: V100
Nvidia Driver Version: 440.64.00
CUDA Version: 10.2
CUDNN Version: 7.6
Operating System + Version: ubuntu18.04
Python Version (if applicable): 3.7
TensorFlow Version (if applicable): None
PyTorch Version (if applicable): 1.6
Baremetal or Container (if container which image + tag): Baremetal

Relevant Files

[06/04/2020-18:07:13] [W] [TRT] onnx2trt_utils.cpp:198: Your ONNX model has been generated with INT64 weights, while TensorRT does not natively support INT64. Attempting to cast down to INT32.
While parsing node number 56 [Einsum]:
ERROR: ModelImporter.cpp:134 In function parseGraph:
[8] No importer registered for op: Einsum
[06/04/2020-18:07:13] [E] Failed to parse onnx file
[06/04/2020-18:07:13] [E] Parsing model failed
[06/04/2020-18:07:13] [E] Engine creation failed
[06/04/2020-18:07:13] [E] Engine set up failed
&&&& FAILED TensorRT.trtexec # /home/juzheng/soft/TensorRT-7.0.0.11/bin/trtexec --onnx=action.onnx --saveEngine=action.trt

2

TensorRT 7.0 supports operators up to Opset 11 and Einsum is only going to be part of the ONNX spec from opset 12.
Please refer below link:

Thanks

3

Thank you for your answer.
So which version of TensorRT will support Einsum.
And is there any alternative？

4

You have to create a custom plugin in TRT to add support for Einsum.

Thanks

5

sorry, I have no ability to implement Einsum function.
I hope you have a simpler solution.

6

You thing you can try is to convert your Pytorch model to Tensorflow model using following flow:
torch → ONNX → .pb
and user TF-TRT API to optimize your model. It will skip the unsupported layer and keep it as tensorflow op.

But for improved performance custom plugin will be the best approach.

Thanks