How to get the tensorRT inference graph based on caffe?

Robotics & Edge Computing Jetson & Embedded Systems Jetson TX2
1

Hi!
I implement R-FCN on TX2 using TensorRT. Some errors occur when I run the .exe file.The error:[Cuda failure: unspecified launch failure at line 74 Aborted (core dumped)]
Now I can’t get the reason of the error. So I want to know how to get the optimization network graph and check it to debug.
Any advice is appreciated.

2

Hi,

.exe is a Windows executable.
Jetson is a Linux system with aarch64 architecture.

Please find the Linux runnable instead.
Thanks.

3

Thank for your reply. Maybe the file’s property is not quite accurate description.The file’s name is inference and its type is executable (application/x-executable) based on aarch64 architecture.
I run the file on the terminal by using the command "./inference and I got some information following by

[Bindings after deserializing:
Binding 0 (data): Input.
Binding 1 (im_info): Input.
Binding 2 (bbox_pred): Output.
Binding 3 (cls_prob): Output.
Binding 4 (rois): Output.
Cuda failure: unspecified launch failure at line 74
Aborted (core dumped)

I found that the error occured when it call the function as following [Specifically this line:context->execute(batchSize, buffers);]
void TensorNet::imageInference(void** buffers, int nbBuffer, int batchSize)
{
assert(engine->getNbBindings()==nbBuffer);

IExecutionContext* context = engine->createExecutionContext();
context->setProfiler(&gProfiler);

context->execute(batchSize, buffers);

context->destroy();

}
I am not sure that the network model after parsing is right because I implement two plugin layers or maybe some other reasons led to the error.
So I want to get the optimization network graph(inference graph) similar to .prototxt file based on caffe so that I can debug it.
Looking forward to your reply!

4

Hi,

Suppose you are using a caffe based model.
Here are the steps to have the layer information:

1. Convert the model into TensorRT network, like this:

const nvcaffeparser1::IBlobNameToTensor* blobNameToTensor = parser->parse(
        locateFile(mParams.prototxtFileName, mParams.dataDirs).c_str(),
        locateFile(mParams.weightsFileName, mParams.dataDirs).c_str(),
        *network,
        nvinfer1::DataType::kFLOAT);

You will have a nvinfer1::INetworkDefinition (network) parameter.

2. Get the total number of layers
https://docs.nvidia.com/deeplearning/sdk/tensorrt-api/c_api/classnvinfer1_1_1_i_network_definition.html#a191a7724fc0c03a3b6f5fd8782dcd30e

virtual int nvinfer1::INetworkDefinition::getNbLayers	(		)	const

3. Check the layer type by the ID:
https://docs.nvidia.com/deeplearning/sdk/tensorrt-api/c_api/classnvinfer1_1_1_i_network_definition.html#a4a81749aaa08e93ca4ae1dbb1739c7bd

virtual ILayer* nvinfer1::INetworkDefinition::getLayer

Thanks.