How does the TensorRT bin file work in PX2?

Hi there,

WE were reading the PX’s object detect sample code and we found that this sample used tensorRT_model.bin file to infer the network. Now we want to integrate a tensorflow model into PX2 and we want to know how to transfer this model into a tensorRT bin file. Is there any instructions? We only found that tensorflow can be transferred into a uff file. Can we use the uff file on px2 as well?

Dear hui_luan,
In DW 0.6, we support only caffe models. We support tensorflow models in next release. We will let you know once schedule is fixed

Thanks for your response Siva!

Is there any instructions about using bin file as TensorRT model? We can’t find any information about that bin file.

Dear hui_luan,
bin file is TensorRT model file. You need to use tensorRT_optimization tool to generate tesnorRT model files which can be used in Driveworks. In the current version, tensorRT_Optmization tool can generate tensorRT models from caffe. We are adding UFF support for tensorRT_Optmization tool in next version.

Hi Siva,

I’m currently using tensorRT_optimization tool to generate tensorRT model from caffe. But it shows an error “could not parse layer type Reshape”. Does that mean it doesn’t support Reshape layer? How can I add that layer into TensorRT model and use it with driveworks?

Dear hui_luan,
unsupported layers can be implemented as custom plugins in TensorRT. But currently, tensorRT_optimization tool in DW does not support custom plugins. We are working on this feature.
So currently, networks which has layers natively supported by TensorRT can be integrated in driveworks.

FYI, reshape layer is supported in TRT 4.0.1 which will be part of next Drive release.

Hi @SivaRamaKrishnaNV , as I am also working with object detect sample code on PX2, I’d like to ask if I need to convert my model into the *.bin format or I can run it in *.uff format. Thanks a lot in advance!

Dear @shd2020,
Yes. You need to get your model file in .bin format using tensorrt_optimization tool to use it with DW DNN framework APIs.