3D convolution in TensorRT

Is it supported in TensorRT 5.0RC? Conversion of a Keras model with 3D convolutions to UFF succeeds, however TensorRT engine construction fails with:
[TensorRT] ERROR: UFFParser: Validator error: conv3d_4/Conv3D: Unsupported operation _Conv3D


Please see https://github.com/NVIDIA-Jetson/redtail/tree/master/stereoDNN for example implementation of conv3d TensorRT plugins.

Is there any plan to integrate what it is in stereoDNN (3D convolution) in the normal model conversion chain of TensorRT?
It isn’t very straightforward the use of code in https://github.com/NVIDIA-Jetson/redtail/tree/master/stereoDNN.
Anyway I think the support of 3D convolution and 3D Max Pooling would be very important for the community, a lot of volume data (Video, Medical Images, etc.) are processed with this type of CNN.


I’m very interested in this too.

Unfortunately the mentioned 3D convolution plugin is not compatible with 3D convolutions from e.g. TensorFlow. TF passes 5D tensors, while TRT appears to throw an assertion. This is even documented in the plugin itself:
https://github.com/NVIDIA-AI-IOT/redtail/blob/master/stereoDNN/lib/conv3d_plugin.cpp#L341 .

The following pull request https://github.com/tensorflow/tensorflow/pull/31461 for TensorFlow hints that TRT 6 will support 3D convolutions with 5D tensors. Can Nvidia tell if this assumption is correct? Are you also planning 3D pooling to be implemented by future versions of TRT?

It is available in tensorrt 7.0, but as said here, few otpimizations are available :/

With TensorRT 7.1.3 the UFF parser is still incapable of loading 3D convolution layers:

UffParser: Validator error: mixed_35x35x256a/branch5x5/Conv/Conv3D: Unsupported operation _Conv3D

Has there been any updates on this ?