Onnx to torchrt convertion error

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

Description

I am trying to convert pytorch model to tensorrt engine.
I had successfully converted pytorch model to onnx but when converting onnx to tensorrt engine it is given below error.

Environment

TensorRT Version: 7.2.1
GPU Type: Jetson tx1
CUDA Version: 10.2

Please include:

[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:119: Searching for input: 2590
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:125: Conv_498 [Conv] inputs: [2000 → (1, 1088, 60, 80)], [2589 → (256, 1088, 1, 1)], [2590 → (256)],
[02/06/2022-19:02:25] [V] [TRT] builtin_op_importers.cpp:450: Convolution input dimensions: (1, 1088, 60, 80)
[02/06/2022-19:02:25] [V] [TRT] ImporterContext.hpp:141: Registering layer: Conv_498 for ONNX node: Conv_498
[02/06/2022-19:02:25] [V] [TRT] builtin_op_importers.cpp:533: Using kernel: (1, 1), strides: (1, 1), prepadding: (0, 0), postpadding: (0, 0), dilations: (1, 1), numOutputs: 256
[02/06/2022-19:02:25] [V] [TRT] builtin_op_importers.cpp:534: Convolution output dimensions: (1, 256, 60, 80)
[02/06/2022-19:02:25] [V] [TRT] ImporterContext.hpp:116: Registering tensor: 2588 for ONNX tensor: 2588
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:179: Conv_498 [Conv] outputs: [2588 → (1, 256, 60, 80)],
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:103: Parsing node: Relu_499 [Relu]
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:119: Searching for input: 2588
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:125: Relu_499 [Relu] inputs: [2588 → (1, 256, 60, 80)],
[02/06/2022-19:02:25] [V] [TRT] ImporterContext.hpp:141: Registering layer: Relu_499 for ONNX node: Relu_499
[02/06/2022-19:02:25] [V] [TRT] ImporterContext.hpp:116: Registering tensor: 2003 for ONNX tensor: 2003
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:179: Relu_499 [Relu] outputs: [2003 → (1, 256, 60, 80)],
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:103: Parsing node: Conv_500 [Conv]
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:119: Searching for input: 2003
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:119: Searching for input: 2592
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:119: Searching for input: 2593
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:125: Conv_500 [Conv] inputs: [2003 → (1, 256, 60, 80)], [2592 → (384, 256, 3, 3)], [2593 → (384)],
[02/06/2022-19:02:25] [V] [TRT] builtin_op_importers.cpp:450: Convolution input dimensions: (1, 256, 60, 80)
[02/06/2022-19:02:25] [V] [TRT] ImporterContext.hpp:141: Registering layer: Conv_500 for ONNX node: Conv_500
[02/06/2022-19:02:25] [V] [TRT] builtin_op_importers.cpp:533: Using kernel: (3, 3), strides: (2, 2), prepadding: (0, 0), postpadding: (0, 0), dilations: (1, 1), numOutputs: 384
[02/06/2022-19:02:25] [V] [TRT] builtin_op_importers.cpp:534: Convolution output dimensions: (1, 384, 29, 39)
[02/06/2022-19:02:25] [V] [TRT] ImporterContext.hpp:116: Registering tensor: 2591 for ONNX tensor: 2591
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:179: Conv_500 [Conv] outputs: [2591 → (1, 384, 29, 39)],
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:103: Parsing node: Relu_501 [Relu]
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:119: Searching for input: 2591
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:125: Relu_501 [Relu] inputs: [2591 → (1, 384, 29, 39)],
[02/06/2022-19:02:25] [V] [TRT] ImporterContext.hpp:141: Registering layer: Relu_501 for ONNX node: Relu_501
[02/06/2022-19:02:25] [V] [TRT] ImporterContext.hpp:116: Registering tensor: 2006 for ONNX tensor: 2006
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:179: Relu_501 [Relu] outputs: [2006 → (1, 384, 29, 39)],
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:103: Parsing node: Conv_502 [Conv]
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:119: Searching for input: 2000
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:119: Searching for input: 2595
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:119: Searching for input: 2596
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:125: Conv_502 [Conv] inputs: [2000 → (1, 1088, 60, 80)], [2595 → (256, 1088, 1, 1)], [2596 → (256)],
[02/06/2022-19:02:25] [V] [TRT] builtin_op_importers.cpp:450: Convolution input dimensions: (1, 1088, 60, 80)
[02/06/2022-19:02:25] [V] [TRT] ImporterContext.hpp:141: Registering layer: Conv_502 for ONNX node: Conv_502
[02/06/2022-19:02:25] [V] [TRT] builtin_op_importers.cpp:533: Using kernel: (1, 1), strides: (1, 1), prepadding: (0, 0), postpadding: (0, 0), dilations: (1, 1), numOutputs: 256
[02/06/2022-19:02:25] [V] [TRT] builtin_op_importers.cpp:534: Convolution output dimensions: (1, 256, 60, 80)
[02/06/2022-19:02:25] [V] [TRT] ImporterContext.hpp:116: Registering tensor: 2594 for ONNX tensor: 2594
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:179: Conv_502 [Conv] outputs: [2594 → (1, 256, 60, 80)],
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:103: Parsing node: Relu_503 [Relu]
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:119: Searching for input: 2594
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:125: Relu_503 [Relu] inputs: [2594 → (1, 256, 60, 80)],
[02/06/2022-19:02:25] [V] [TRT] ImporterContext.hpp:141: Registering layer: Relu_503 for ONNX node: Relu_503
[02/06/2022-19:02:25] [V] [TRT] ImporterContext.hpp:116: Registering tensor: 2009 for ONNX tensor: 2009
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:179: Relu_503 [Relu] outputs: [2009 → (1, 256, 60, 80)],
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:103: Parsing node: Conv_504 [Conv]
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:119: Searching for input: 2009
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:119: Searching for input: 2598
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:119: Searching for input: 2599
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:125: Conv_504 [Conv] inputs: [2009 → (1, 256, 60, 80)], [2598 → (288, 256, 3, 3)], [2599 → (288)],
[02/06/2022-19:02:25] [V] [TRT] builtin_op_importers.cpp:450: Convolution input dimensions: (1, 256, 60, 80)
[02/06/2022-19:02:25] [V] [TRT] ImporterContext.hpp:141: Registering layer: Conv_504 for ONNX node: Conv_504
[02/06/2022-19:02:25] [V] [TRT] builtin_op_importers.cpp:533: Using kernel: (3, 3), strides: (2, 2), prepadding: (0, 0), postpadding: (0, 0), dilations: (1, 1), numOutputs: 288
[02/06/2022-19:02:25] [V] [TRT] builtin_op_importers.cpp:534: Convolution output dimensions: (1, 288, 29, 39)
[02/06/2022-19:02:25] [V] [TRT] ImporterContext.hpp:116: Registering tensor: 2597 for ONNX tensor: 2597
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:179: Conv_504 [Conv] outputs: [2597 → (1, 288, 29, 39)],
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:103: Parsing node: Relu_505 [Relu]
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:119: Searching for input: 2597
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:125: Relu_505 [Relu] inputs: [2597 → (1, 288, 29, 39)],
[02/06/2022-19:02:25] [V] [TRT] ImporterContext.hpp:141: Registering layer: Relu_505 for ONNX node: Relu_505
[02/06/2022-19:02:25] [V] [TRT] ImporterContext.hpp:116: Registering tensor: 2012 for ONNX tensor: 2012
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:179: Relu_505 [Relu] outputs: [2012 → (1, 288, 29, 39)],
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:103: Parsing node: Conv_506 [Conv]
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:119: Searching for input: 2000
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:119: Searching for input: 2601
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:119: Searching for input: 2602
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:125: Conv_506 [Conv] inputs: [2000 → (1, 1088, 60, 80)], [2601 → (256, 1088, 1, 1)], [2602 → (256)],
[02/06/2022-19:02:25] [V] [TRT] builtin_op_importers.cpp:450: Convolution input dimensions: (1, 1088, 60, 80)
[02/06/2022-19:02:25] [V] [TRT] ImporterContext.hpp:141: Registering layer: Conv_506 for ONNX node: Conv_506
[02/06/2022-19:02:25] [V] [TRT] builtin_op_importers.cpp:533: Using kernel: (1, 1), strides: (1, 1), prepadding: (0, 0), postpadding: (0, 0), dilations: (1, 1), numOutputs: 256
[02/06/2022-19:02:25] [V] [TRT] builtin_op_importers.cpp:534: Convolution output dimensions: (1, 256, 60, 80)
[02/06/2022-19:02:25] [V] [TRT] ImporterContext.hpp:116: Registering tensor: 2600 for ONNX tensor: 2600
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:179: Conv_506 [Conv] outputs: [2600 → (1, 256, 60, 80)],
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:103: Parsing node: Relu_507 [Relu]
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:119: Searching for input: 2600
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:125: Relu_507 [Relu] inputs: [2600 → (1, 256, 60, 80)],
[02/06/2022-19:02:25] [V] [TRT] ImporterContext.hpp:141: Registering layer: Relu_507 for ONNX node: Relu_507
[02/06/2022-19:02:25] [V] [TRT] ImporterContext.hpp:116: Registering tensor: 2015 for ONNX tensor: 2015
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:179: Relu_507 [Relu] outputs: [2015 → (1, 256, 60, 80)],
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:103: Parsing node: Conv_508 [Conv]
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:119: Searching for input: 2015
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:119: Searching for input: 2604
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:119: Searching for input: 2605
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:125: Conv_508 [Conv] inputs: [2015 → (1, 256, 60, 80)], [2604 → (288, 256, 3, 3)], [2605 → (288)],
[02/06/2022-19:02:25] [V] [TRT] builtin_op_importers.cpp:450: Convolution input dimensions: (1, 256, 60, 80)
[02/06/2022-19:02:25] [V] [TRT] ImporterContext.hpp:141: Registering layer: Conv_508 for ONNX node: Conv_508
[02/06/2022-19:02:25] [V] [TRT] builtin_op_importers.cpp:533: Using kernel: (3, 3), strides: (1, 1), prepadding: (1, 1), postpadding: (1, 1), dilations: (1, 1), numOutputs: 288
[02/06/2022-19:02:25] [V] [TRT] builtin_op_importers.cpp:534: Convolution output dimensions: (1, 288, 60, 80)
[02/06/2022-19:02:25] [V] [TRT] ImporterContext.hpp:116: Registering tensor: 2603 for ONNX tensor: 2603
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:179: Conv_508 [Conv] outputs: [2603 → (1, 288, 60, 80)],
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:103: Parsing node: Relu_509 [Relu]
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:119: Searching for input: 2603
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:125: Relu_509 [Relu] inputs: [2603 → (1, 288, 60, 80)],
[02/06/2022-19:02:25] [V] [TRT] ImporterContext.hpp:141: Registering layer: Relu_509 for ONNX node: Relu_509
[02/06/2022-19:02:25] [V] [TRT] ImporterContext.hpp:116: Registering tensor: 2018 for ONNX tensor: 2018
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:179: Relu_509 [Relu] outputs: [2018 → (1, 288, 60, 80)],
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:103: Parsing node: Conv_510 [Conv]
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:119: Searching for input: 2018
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:119: Searching for input: 2607
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:119: Searching for input: 2608
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:125: Conv_510 [Conv] inputs: [2018 → (1, 288, 60, 80)], [2607 → (320, 288, 3, 3)], [2608 → (320)],
[02/06/2022-19:02:25] [V] [TRT] builtin_op_importers.cpp:450: Convolution input dimensions: (1, 288, 60, 80)
[02/06/2022-19:02:25] [V] [TRT] ImporterContext.hpp:141: Registering layer: Conv_510 for ONNX node: Conv_510
[02/06/2022-19:02:25] [V] [TRT] builtin_op_importers.cpp:533: Using kernel: (3, 3), strides: (2, 2), prepadding: (0, 0), postpadding: (0, 0), dilations: (1, 1), numOutputs: 320
[02/06/2022-19:02:25] [V] [TRT] builtin_op_importers.cpp:534: Convolution output dimensions: (1, 320, 29, 39)
[02/06/2022-19:02:25] [V] [TRT] ImporterContext.hpp:116: Registering tensor: 2606 for ONNX tensor: 2606
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:179: Conv_510 [Conv] outputs: [2606 → (1, 320, 29, 39)],
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:103: Parsing node: Relu_511 [Relu]
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:119: Searching for input: 2606
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:125: Relu_511 [Relu] inputs: [2606 → (1, 320, 29, 39)],
[02/06/2022-19:02:25] [V] [TRT] ImporterContext.hpp:141: Registering layer: Relu_511 for ONNX node: Relu_511
[02/06/2022-19:02:25] [V] [TRT] ImporterContext.hpp:116: Registering tensor: 2021 for ONNX tensor: 2021
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:179: Relu_511 [Relu] outputs: [2021 → (1, 320, 29, 39)],
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:103: Parsing node: MaxPool_512 [MaxPool]
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:119: Searching for input: 2000
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:125: MaxPool_512 [MaxPool] inputs: [2000 → (1, 1088, 60, 80)],
[02/06/2022-19:02:25] [V] [TRT] ImporterContext.hpp:141: Registering layer: MaxPool_512 for ONNX node: MaxPool_512
[02/06/2022-19:02:25] [V] [TRT] ImporterContext.hpp:116: Registering tensor: 2022 for ONNX tensor: 2022
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:179: MaxPool_512 [MaxPool] outputs: [2022 → (1, 1088, 29, 39)],
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:103: Parsing node: Concat_513 [Concat]
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:119: Searching for input: 2006
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:119: Searching for input: 2012
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:119: Searching for input: 2021
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:119: Searching for input: 2022
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:125: Concat_513 [Concat] inputs: [2006 → (1, 384, 29, 39)], [2012 → (1, 288, 29, 39)], [2021 → (1, 320, 29, 39)], [2022 → (1, 1088, 29, 39)],
[02/06/2022-19:02:25] [V] [TRT] ImporterContext.hpp:141: Registering layer: Concat_513 for ONNX node: Concat_513
[02/06/2022-19:02:25] [V] [TRT] ImporterContext.hpp:116: Registering tensor: 2023 for ONNX tensor: 2023
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:179: Concat_513 [Concat] outputs: [2023 → (1, 2080, 29, 39)],
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:103: Parsing node: Conv_514 [Conv]
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:119: Searching for input: 2023
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:119: Searching for input: fpn.lateral4.weight
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:125: Conv_514 [Conv] inputs: [2023 → (1, 2080, 29, 39)], [fpn.lateral4.weight → (256, 2080, 1, 1)],
[02/06/2022-19:02:25] [V] [TRT] builtin_op_importers.cpp:450: Convolution input dimensions: (1, 2080, 29, 39)
[02/06/2022-19:02:25] [V] [TRT] ImporterContext.hpp:141: Registering layer: Conv_514 for ONNX node: Conv_514
[02/06/2022-19:02:25] [V] [TRT] builtin_op_importers.cpp:533: Using kernel: (1, 1), strides: (1, 1), prepadding: (0, 0), postpadding: (0, 0), dilations: (1, 1), numOutputs: 256
[02/06/2022-19:02:25] [V] [TRT] builtin_op_importers.cpp:534: Convolution output dimensions: (1, 256, 29, 39)
[02/06/2022-19:02:25] [V] [TRT] ImporterContext.hpp:116: Registering tensor: 2024 for ONNX tensor: 2024
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:179: Conv_514 [Conv] outputs: [2024 → (1, 256, 29, 39)],
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:103: Parsing node: Pad_515 [Pad]
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:119: Searching for input: 2024
[02/06/2022-19:02:25] [V] [TRT] ModelImporter.cpp:125: Pad_515 [Pad] inputs: [2024 → (1, 256, 29, 39)],
ERROR: builtin_op_importers.cpp:2191 In function importPad:
[8] Assertion failed: mode == “constant” && value == 0.f && “This version of TensorRT only supports constant 0 padding!”
[02/06/2022-19:02:25] [E] Failed to parse onnx file
[02/06/2022-19:02:25] [E] Parsing model failed
[02/06/2022-19:02:25] [E] Engine creation failed
[02/06/2022-19:02:25] [E] Engine set up failed
&&&& FAILED TensorRT.trtexec # /usr/src/tensorrt/bin/trtexec --onnx=deblur.onnx --verbose

2

Hi,

Looks like you’re using an old version of TensorRT. I believe this issue has been fixed in the new release. We recommend you to please try on latest TensorRT verison 8.2
Please refer following similar issue for more details.
https://github.com/NVIDIA/TensorRT/issues/1019

Thank you.

3

I can not use latest version due to business reason. Can I do any other workaround with model or add that layer in trt 7.

4

Hi,

Unfortunately there is no workaround, if we are using non-0 or non-constant padding TRT 7 cannot handle it.

Thank you.