Onnx-tensorrt: WARNING: ONNX model has a newer ir_version (0.0.5) than this parser was built against (0.0.3).

[electronicYH]

I meet the problem:

a@a:/data/yh/onnx-tensorrt/build$ onnx2trt yh_fcn_14c_include_weights.onnx -o yh.trt -g

Input filename: yh_fcn_14c_include_weights.onnx
ONNX IR version: 0.0.5
Opset version: 10
Producer name: keras2onnx
Producer version: 1.5.0
Domain: onnx
Model version: 0

Doc string:

WARNING: ONNX model has a newer ir_version (0.0.5) than this parser was built against (0.0.3).
Parsing model
While parsing node number 1 [Conv -> "convolution_output"]:
ERROR: /data/yh/onnx-tensorrt/ModelImporter.cpp:537 In function importModel:
[5] Assertion failed: tensors.count(input_name)

please help me to fix it, thank you so much!!!!!

[he44]

Hi @electronicYH did you fix the issue? I had the same problem here.
I tried to resolve the ir version warning by setting different target opset in keras2onnx but that didn't help. (I tried all the way form 1 to 10).
For the assertion failure, I checked the ModelImport.cpp. I assume it's because the code finds that one of the nodes have 0 input? Though I checked the graph of my model, don't think that's the case.

for( size_t node_idx : topological_order ) {
    _current_node = node_idx;
    ::ONNX_NAMESPACE::NodeProto const& node = graph.node(node_idx);
    std::vector<TensorOrWeights> inputs;
    for( auto const& input_name : node.input() ) {
      ASSERT(tensors.count(input_name), ErrorCode::kINVALID_GRAPH);
      inputs.push_back(tensors.at(input_name));
    }
    // ...
}

On a separate note, have you run the onnx checker after exporting?

import onnx
onnx.checker.check_model(onnx_model)

The checker threw a warning for me:

Nodes in a graph must be topologically sorted, however input 'conv2d_31_Relu_0' of node: 
input: "conv2d_31_Relu_0" output: "transpose_output7" name: "Transpose21" op_type: "Transpose" attribute { name: "perm" ints: 0 ints: 2 ints: 3 ints: 1 type: INTS } doc_string: "" domain: ""
 is not output of any previous nodes.

However, my onnx model can run inference fine in Python API using the code below. So I assumed this shouldn't be a big issue?

import onnxruntime as rt
import numpy as np
X_test = np.ones((1,h,w,1))
sess    = rt.InferenceSession("my_model.onnx")
input_name = sess.get_inputs()[0].name
label_name = sess.get_outputs()[0].name
pred_onx    = sess.run([label_name], {input_name: X_test.astype(np.float32)})[0]

Any insights would be greatly appreciated!

[pango99]

I ran into the same problem when I ran the DeepLabV3 onnx model converted from keras on TensorRT SDK 5.1.5, how to fix the problem?

[zjd1988]

@pango99 I modify importModel code in ModelImporter.cpp, and rcompile onnx-tensorrt(https://github.com/onnx/onnx-tensorrt v5.0)

it works for me, you can have a try.
ModelImporter.cpp.zip

[pango99]

• [ ]

@pango99 I modify importModel code in ModelImporter.cpp, and rcompile onnx-tensorrt(https://github.com/onnx/onnx-tensorrt v5.0)

it works for me, you can have a try.
ModelImporter.cpp.zip

Thanks zjd1988,so can I compile onnx-tensorrt on windows platform? and can I use the compiled module with nvidia tensorrt sdk directly?

[zjd1988]

I compile onnx-tensorrt on ubuntu16.04, but it might be works well on windows, refferd from https://github.com/onnx/onnx-tensorrt/issues/60 .

you need replace the old libs (in tensorrt sdk) with newly compiled, or modify the LD_LIBRARY_PATH(add the path of onnx-tensorrt libs)

[pango99]

I compile onnx-tensorrt on ubuntu16.04, but it might be works well on windows, refferd from #60 .

you need replace the old libs (in tensorrt sdk) with newly compiled, or modify the LD_LIBRARY_PATH(add the path of onnx-tensorrt libs)

OK,I'll try to compile few days later

[drormeir]

@pango99 I modify importModel code in ModelImporter.cpp, and rcompile onnx-tensorrt(https://github.com/onnx/onnx-tensorrt v5.0)

it works for me, you can have a try.
ModelImporter.cpp.zip

I used your source code and now I get different assert error:

While parsing node number 0 [Slice]:
3
ERROR: /home/ubuntu/projects/onnx-tensorrt/builtin_op_importers.cpp:1687 In function importSlice:
[6] Assertion failed: inputs.size() == 1

Any help would be appreciated.

[zjd1988]

Hi ＠drormeir .You can debug the onnx-tensorrt code step by step, to find out which layer assert.

After that,you can check the onnx-tensorrt suported ops from this link https://github.com/onnx/onnx-tensorrt/blob/master/operators.md

[ttdd11]

@pango99 Same issue with the parser. Ran onnx.checker.check_model(onnx_model) with no error. Were you able to build these libs on windows and resolve this error?

[j-o-d-o]

As @ttdd11 said, get same error while the checker did not complain.

[j-o-d-o]

Tried using the changes from @drormeir but somehow non of my changes actually have an effect. Also the line number of the assert does not match the one of the ModelImporter.cpp I have.
That's what my cmake looks like:

add_library(libnvonnxparser SHARED IMPORTED)
set_property(TARGET libnvonnxparser PROPERTY IMPORTED_LOCATION ${INSTALL_DIR}/onnx-tensorrt/dist/lib/libnvonnxparser.so)
add_library(libnvinfer SHARED IMPORTED)
set_property(TARGET libnvinfer PROPERTY IMPORTED_LOCATION $ENV{TRT_RELEASE}/lib/libnvinfer.so)
...
target_link_libraries(${PROJECT_NAME}
  ...
  libnvonnxparser
  libnvinfer
}

It seems that another ModelImporter.cpp is used than the one I am building but the one from the TENSORRT_ROOT folder :/ Anyone else had the same issue?

[j-o-d-o]

Updating to the recent 7.0.0 release of TensorRT (including 7.0.0 onnx-tensorrt) fixed the whole thing. Both the assertion and tensorrt using and old onnx-tensorrt version. At least in my case.

[tcxia]

• [ ]

@pango99 I modify importModel code in ModelImporter.cpp, and rcompile onnx-tensorrt(https://github.com/onnx/onnx-tensorrt v5.0)

it works for me, you can have a try.
ModelImporter.cpp.zip

Thanks zjd1988,so can I compile onnx-tensorrt on windows platform? and can I use the compiled module with nvidia tensorrt sdk directly?

i copy your code and get error : ‘initializer_map’ was not declared in this scope

[zjd1988]

Hi @tcxia, which tensorrt version you use?

[tcxia]

Hi @tcxia, which tensorrt version you use?

tensorRT 5.1 cuda10.1

[zjd1988]

Hi @tcxia, I tested ModelImporter.cpp.zip with tensorRT5.0, I not sure ablout tensorRT5.1. could you try this with tensorRT5.0.

[Toku11]

I was facing same issue in tensor docker container, solution was sh /opt/tensorrt/install_opensource.sh

[kevinch-nv]

This issue was fixed in later TensorRT versions. If you are still having trouble with the current release version (7.1) please open another issue.

[austinmw]

@kevinch-nv Hi, I'm unable to upgrade my TRT version as I need to use CUDA 10.2. Is there any way to solve this issue with TensorRT 6.0.1? Or is it possible to upgrade beyond 6.0.1 with CUDA 10.2?