Cannot build the pb model (tensorflow), got "tf_graph_parser.cpp(48): Creating input desc is failed." .

[tencent]

Cannot build the pb model (tensorflow), got "tf_graph_parser.cpp(48): Creating input desc is failed." .

[yuanzexi]

1. For you first problem, the inputs you used to build engine should be the same as the inputs of your frozen models. I found that you used (x1: 10, x2: 3) as the inputs of the frozen models, but [1, 768] as the inputs to build engine. They are not matched such that Creating input desc is failed.
2. The problem about invalid pointer might be related to conflicts with TensorRT libraries on BERT-related plugins registry. We'll try to reproduce it to figure out.

[favorxin]

Thanks for answering~

1. For the first problem, I check the input and solved the input problem. But another problem (已放弃(吐核)) can not solve, I load the pb model to build the engine and the problem appeared.

[yuanzexi]

@favorxin Did you try to pull the latest master branch to rebuild? Because we had fixed a Plugin Registry problem based on your previous 'unit_test' screenshot, you could pull the latest master branch and rebuild it again.

[favorxin]

Sorry for replaying later. I tried again yesterday and pulled the latest master branch. I have a question that the tensorflow==1.15.0 is tensorflow-gpu? But my cuda=11.0, it too higher for tensorflow=1.15.0. It got libcudnn.so.7 not found and core dumped wrong again.

[yuanzexi]

@favorxin Maybe you could try to config your ENV variable. Some references are below:

export PATH=/usr/local/cuda/bin:$PATH
export LD_LIBRARY_PATH=/usr/lib/:/usr/lib64/:/usr/local/cuda/lib64/:/usr/local/nvidia/lib64:/usr/local/cuda/extras/CUPTI/lib64/:$LD_LIBRARY_PATH
export LIBRARY_PATH=/usr/lib:/usr/lib64/:/usr/local/cuda/lib64/:$LIBRARY_PATH
export CPLUS_INCLUDE_PATH=/usr/local/cuda/include/:$CPLUS_INCLUDE_PATH

[favorxin]

Here is my keras model.
densenet_num18_all_model.zip

My cmake code is as following:
cmake .. -DTensorRT_ROOT=/data/software/tensorRT/TensorRT-7.0.0.11 \ -DENABLE_LOGGING=OFF \ -DENABLE _PROFILING=OFF \ -DENABLE_DYNAMIC_BATCH=ON \ -DBUILD_PYTHON_LIB=ON \ -DPYTHON_EXECUTABLE=/data/anaconda3/bin/python \ -DENABLE_TORCH=OFF \ -DENABLE_TENSORFLOW=OFF \ -DENABLE_KERAS=ON \ -DCMAKE_PREFIX_PATH=/data/software/hdf5-1.12.0
And this log is cmake log.

The make -j is successfully.

Environment

TensorRT Version: 7.0.0.6
NVIDIA GPU: K40S
NVIDIA Driver Version: 450.80.02
CUDA Version: 10.0
CUDNN Version: 7.6.5
Operating System: 7.5
Python Version (if applicable): 3.6.13
Tensorflow Version (if applicable): tensorflow==1.15.0(cpu)
PyTorch Version (if applicable): 1.7.1

[github-actions]

This issue is stale because it has been open 30 days with no activity. Remove stale label or comment or this will be closed in 5 days.

[github-actions]

This issue was closed because it has been stalled for 5 days with no activity.