UFFParser: Validator error: Unsupported operation _Identity

Accelerated Computing Intelligent Video Analytics DeepStream SDK
1

I converted a Tensorflow frozen model based on the SSD300(https://github.com/oarriaga/SSD-keras) to an UFF on NVIDIA TensorRT 3.0.2 library.

I am met with a UFFParser: Validator error: SSD300_20: Unsupported operation _Identity when I call IDeviceWorker::addInferenceTask_uff method in my program based on DeepStream SDK 1.5.
(SSD300_20 is an output layer name of the Tensorflow frozen model above.)

My questions are

  1. Is this UFFParser error caused by calling tf.identity in Tensorflow API or a centain Keras API?
  2. How can I fix this UFFParser error?
1 Like
2

Hi,

Identity layer is not supported by TensorRT.
The error shows unsupported information when creating TensorRT engine.

In our experience, the Identity layer is not really useful but it’s added by Keras when bypassing the dropout during inference.
Please try to refine your model without Identity since it can be removed safely.

Thanks.

3

I think this is needed. In my use case, I have several .protobuf layers which has the same name as prefix. For example vlas_layer_1/Square, vlas_layer1/Sqaa, vlas_layer1/Normed, vlas_layer1. Since several layers have the same prefix, UFF parser is confused. So I had set vlas_layer1 through the identity operation, effectively renaming it.

What should I do in my case?