yolov3_onnx, onnx_to_tensorrt.py <tensorrt.tensorrt.Builder object at 0x7f5d678a78>, 0 bug about tensorrt_demos

[tnl-mid]

I had the same issue, fixed it by replacing
builder.create_network(EXPLICIT_BATCH)
with:
builder.create_network()

.. in onnx_to_tensorrt.py. Detection works fine now. The EXPLICIT_BATCH argument is specific for TensorRT7, I think.
I also had the message for 'Layer of type yolo not supported', but it all seemed to work nonetheless.

[jkjung-avt]

Sorry, this is my bad. I modified the code to support TensorRT 7 (for x86_64) a couple of days ago. I did not do enough testing to make sure the modification does not break TensorRT 5 & 6. I just pushed a fix into the repository.

Thanks to Michael-Dub for quick answering of the question.

[omeralierdemir]

thanks your help. Michael-Dub advise worked. I have few quastions ? My system tensorRT 6.0 on jetson tx2 and I try to convert my onnx_model to tensorRT C++ API but I couldn't. Do you have any idea for this quastions? I tried this repo https://github.com/onnx/onnx-tensorrt but tx2 not suporrted tensorRT 7.0. https://devtalk.nvidia.com/default/topic/1069634/how-to-upgrade-tensorrt-to-7-0-version/?offset=7

[jkjung-avt]

@omeralierdemir Would you like to inference (do object detection) with TensorRT C++ API, or to convert a YOLOv3 model to TensorRT engine with C++? If it's the former case, you could use the TensorRT engine created by my python scripts, and:

• Reference my TrtGooglenet code for how to feed an input image ('float32' array) and do inference with the TensorRT engine.
• Implement the preprocessing and postprocessing code in C++.

I think a C++ implementation could run significantly faster than python if done properly.