ArgMax/ExpandDims/Slice operations in NVIDIA DRIVE TRT4 for Tensorflow/Caffe

Hi All,

I would like any suggestion on ArgMax/ExpandDims/Slice operations in NVIDIA DRIVE TensorRT4 for Tensorflow/Caffe via customized layer. Any advice is always helpful!

Thanks!

Gary

Hi,

These operations will be supported in our next release but not available currently.
As an alternative, you can implement it with plugin API.

Since TensorRT 5.0 is available for the DRIVE AGX user now, is it an option for you?
You can get a new plugin API in TensorRT 5.0, which is much easier to implement.

Thanks.

May I get any early version of tensorrt 5 for AGX DRIVE (preparing for Pegasus via DRIVE early access program)? Thanks!

Hi,

Sorry that there is something incorrect in my previous comment.

Here is an example of plugin layer implementation with TensorRT4.0:
https://docs.nvidia.com/deeplearning/sdk/tensorrt-archived/tensorrt_401/tensorrt-developer-guide/index.html#uffssd_sample

Since you will need a desktop environment to convert the .pb file into uff, it’s recommended to do this implementation on the desktop first.
After getting a correct result, you can reproduce it on DRIVE AGX with the TensorRT c++ code and uff model.

Thanks.

Hi, garywang

Sorry that some message in comment#2 is incorrect (already make as strikethrough).

TensorRT 5.0 is not available for DRIVE platform yet.
Please follow the tutorial shared in comment#4 to enable the plugin layer with TensorRT4.0.

Thanks.