tensorRT_optimization for semantic classification

Autonomous Machines DRIVE PX2 DriveWorks
#1

Hi,

I trained a semantic classification network using digits 5.0 and want to test it on the DPX2.
Therefore, I wanted to create the optimized version using tensorRT_optimization tool, and I got the following error:
./tensorRT_optimization --prototxt=deployVOC.prototxt --caffemodel=weightsVOC.caffemodel --outputBlobs=score --modelType=caffe
Initializing network optimizer on model deployVOC.prototxt with weights from weightsVOC.caffemodel
could not parse layer type Crop
Error: DW_INVALID_ARGUMENT: Provided network could not be parsed.

Is it possible to use semantic classification on DPX2? What did I wrong?

Thanks for your support.

Best,
Sacha

#2

Dear sacha.vrazic,

Could you please re-try it with below command in /usr/src/tensorrt/bin on DPX2?

./giexc --deploy=[deploy prototxt] --model=[caffemodel file] --output=[output tensor] –verbose –hostTime –int8(or –half2, or nothing).

#3

Dear Steve,

I tried but unfortunately the error is the same…
Any other suggestion?

#4

Dear sacha.vrazic,

May I know what TensorRT version is?
If SDK version is 5.0.5.0, TensorRT version is 3.0RC.
And please refer to below webinar link for TensorRT. Thanks.
http://info.nvidia.com/drivepx-tensorrt-reg-page.html

#5

Dear Steve,

Indeed, the SDK version is 5.0.5.0.
I checked the webinar, and the tensorRT_optimization file is not used, and some python code is shown.
Therefore, what is for the tensorRT_optimization file, and do we need to install tensorRT 3.0 on our host?

Could you send the files that are used in the webinar (the fcn and the python files)?
Thank you in advance.

Best,
Sacha

#6

Dear sacha.vrazic,

I found the symptom reason, it’s because TensorRT doesn’t support Crop layer from Caffe.
It will need to be implemented using custom layer (plugin).

Also, DriveWorks tool ‘tensorRT_optimization’ doesn’t support custom layer, so if you use any network that requires custom layer in TensorRT, you cannot use DriveWorks API for DNNs. Thanks.