cudnn6 slow and problematic on TX2, JetPack 3.1

Autonomous Machines Jetson & Embedded Systems Jetson TX2
21

Yes, I want to concatenate multiple tensors along the channel dimension in fp16.

All the concatenate layers in the prototxt I linked perform the operation on axis=1 which in Caffe is the channel axis (format is NCHW), so even if the parameter gets ignored it should work.

22

Hi,

There are some fixes in TensorRT3.0 package.
Could you check tensorRT3.0 package first?

Thanks.