I would like to if tensorflow spaceToDepth functionality is supported with movidius ncsdk 2.10 ?

[AAhma26]

I am compiling YOLO v2 network and I am facing following error:

[Error 4] Toolkit Error: Stage Type Not Supported: SpaceToDepth.

Please let me know if there are alternative way for this.

Thanks

[JesusE_Intel]

Hi AAhma26,

The Intel Movidius Neural Compute SDK does not support the Tensorflow SpaceToDepth, you can find a list of supported networks and layers on the release notes. However, there is a GitHub project from a user who has instructions on running tiny yolo v2 with the Intel Movidius NCSDK. I have not tried it myself but give it a try and see if it helps.

Did you custom trained your YOLOv2 or are you using a pre-trained network? Could you share a link to your model? Is there a particular reason why you are using YOLOv2 vs YOLOv3?

Regards,

Jesus

[AAhma26]

Hi,

Thanks for your response.

I am working on my Final year Project and I need to optimize weights and layers which could not be easily done with the darkflow Tiny yolo v2. Therefore, I had to implement YOLO v2 from scratch. I am completely flexible with changing the network to the yolo network supported by movidius. If you can give me any link to the network architecture or an alternate function for space to depth, it will be great.

Another question is can I implement my own space to depth as space to depth function is just reshaping and swapping axis, will that be a problem in compiling with Movidius?

Regards

[JesusE_Intel · Accepted solution]

Hi AAhma26,

The Intel Movidius Neural Compute SDK does not support creating custom layers so you won't be able to implement it. The recent release of the OpenVINO toolkit has added support for TensorFlow SpaceToDepth operation. I recommend trying to convert your model using the OpenVINO toolkit.

Take a look at the following links/documentation:

Getting Started Guide for Intel NCS 2 (Same steps for NCS 1)

Download page for OpenVINO toolkit

Converting TensorFlow Models

Hope this helps!

Regards,

Jesus