I am getting the error:
[Error 5] Toolkit Error: Stage Details Not Supported: Transpose
Is there any way around this? Is there a reason why Transpose is not implemented? If not, is their a time-frame for seeing the functionality added to the NCS?
@peterf Thanks for reporting this. Can you provide a link to the network for testing? When adding initial TensorFlow support, we focused on a handful of networks and added the layers needed to run those networks. There are quite a few of TensorFlow ops and this one is likely one that hasn't been implemented yet. We can't provide a time frame for adding this op at the moment.
If you are using transpose for input manipulation like changing the input from HxWxC to CxHxW, you can use cut this op out of the model and use numpy's transpose in your python code before making an inference on the NCS. FYI the NCSDK likes HxWxC input.
@Tome_at_Intel , I actually cannot link the network for testing, but I can run any tests you'd like for me to run. I would prefer to not have to use numpy for transposing, due to my current workflow:
Build network in PyTorch -> Compile to ONNX -> ONNX to TF -> TF to NCSK.
PyTorch's native convolution ops utilize transposing, and therein lies the problem.