Generate TensorRT from Caffemodel Error

Autonomous Vehicles DRIVE PX 2 DriveWorks
1

Hi,
A problem occurred when I transfered my own Caffemodel to the TensorRT model.
LSTM modules are added in my caffemodel, and it failed when I used the driveworks API to generate TensorRT model from my Caffemodel.

The error is like:
Error parsing text-format ditcaffe.NetParameter: 15:19: Message type “ditcaffe.LayerParameter” has no field named “recurrent_param”.)

It seems like TensorRT can not parse the LSTM layer in Caffe model. However, in TensorRT developer guide, I find that the LSTM layer is supported by tensorRT4.

The version of driveworks on my ubuntu16.04 is 1.2.
Please let me know if you need any other info.

Thanks
Issac

2

Dear wh120355,
We are looking into the issue. We will update you.

3

Dear wh120355,
Unfortunately we don’t support LSTM layer in caffe model currently.

4

I also have the same problem on parsing the lstm layer of caffe model.
Is there any solutions to solve this unsupport problem? Wait till you suport or code by myself?

5

From TensorRT perspective, it can be done by implementing a custom layer. You can take a look at [url]https://docs.nvidia.com/deeplearning/sdk/tensorrt-developer-guide/index.html#add_custom_layer[/url].

But for DriveWorks, We focus on the models we develop. And to enable a plugin layer may need to compile the custom code into library, which is not available for currently DWs.