Tensorrt: How can I use tensorRT in 3D image including hight, width and depth

Created on 13 Aug 2019 · 3Comments · Source: NVIDIA/TensorRT

Hi ,
I have a tensorflow-trained 3D network---V-Net, which is a 3D segmentation network. That means the input and output image is a 3D image including hight, width and depth. I notice that the tensorRT API UffInputOrder has three members:NC, NHWC, NCHW. All of them are for 2D image.
So how can I use tensorRT in my 3D V-Net model?
Thanks

Source

AlphaJia

Most helpful comment

Hi ,
I have a tensorflow-trained 3D network---V-Net, which is a 3D segmentation network. That means the input and output image is a 3D image including hight, width and depth. I notice that the tensorRT API UffInputOrder has three members:NC, NHWC, NCHW. All of them are for 2D image.
So how can I use tensorRT in my 3D V-Net model?
Thanks

Now TensorRT 6.0 supports conv3D.

Rongnian on 17 Sep 2019

👍2 🎉1

All 3 comments

TensoRT only supports 2D convolution.
https://docs.nvidia.com/deeplearning/sdk/tensorrt-developer-guide/index.html#convolution-layer

Rongnian on 14 Aug 2019

Hi ,
I have a tensorflow-trained 3D network---V-Net, which is a 3D segmentation network. That means the input and output image is a 3D image including hight, width and depth. I notice that the tensorRT API UffInputOrder has three members:NC, NHWC, NCHW. All of them are for 2D image.
So how can I use tensorRT in my 3D V-Net model?
Thanks

Now TensorRT 6.0 supports conv3D.

Rongnian on 17 Sep 2019

👍2 🎉1

@Rongnian is correct. And if you experience issues with the UFF parser, I would recommend exporting your model to ONNX and trying the ONNX parser instead like here: https://github.com/NVIDIA/TensorRT/issues/185.

If you experience this error (https://github.com/NVIDIA/TensorRT/issues/180), it should be fixed in a future release.