Tensorflow supported operations question



I’m working on autonomous cars in a university setting and we would like to create a lane detection model in Tensorflow (1.15.0 - Python API) that is compatible with the native TensorRT API so we can create an optimized C++ inference engine. I checked the support matrix you provided for the TensorRT version we use ( here to see what TF operations are supported:


My questions would be:
1 .How do I know what TensorFlow module these listed operations belong to? Do I have to use the tf.nn module or perhaps is it okay to use the tf.keras module (we were working with this) ?

  1. Why Dense layer is not listed in the supported operations? I’m asking because it is quite common to use it in deep networks one way or another and I can’t see how can I create the TF network without it.

Thanks in advance!


TensorRT Version:
GPU Type: GeForce RTX 2080 Ti (3x)
Nvidia Driver Version: 450.51.05
CUDA Version: 10.0
CUDNN Version: 7.6.5
Operating System + Version: Ubuntu 18.04
Python Version (if applicable): 3.6.9
TensorFlow Version (if applicable): 1.15.0
PyTorch Version (if applicable): -
Baremetal or Container (if container which image + tag): -

Hi @norbertmarko92,

We recommend you to use the latest TRT version(7.1) with ONNX flow for better op support.
You can find it here

However for UFF supported ops, you can check the below link

For unsupported ops, you need to create custom layers.
Check the below link for reference

1 Like

Thank you for your response! I will consider the ONNX workflow for the future. However I do not understand how a layer like Dense could be missing from the list of supported operations since it is everywhere in neural networks. Am I missing something or is it listed under a different name?

Hi @norbertmarko92,
We have fullyConnected Layer in TRT which is named as Dense in other network.