0

This issue relates to IR inferencing using OpenVINO on NCS2 (MYRIAD). The ONNX representation of the model I am trying to convert comprises 8 ScatterNDupdate layers, which OpenVINO doesn't support on NCS2.

  1. Is there any way to get rid of those layers without affecting the model's functionality?
  2. Does OpenVINO provide any alternative layer which will not affect the inferences quantitatively or qualitatively?
  3. For future references, is there any way to avoid of scatterNDupdate layers? (reason for the creation of these layers)

Link to IR and model

Details:

OpenVINO version: 2022.1

ONNX Opset: 11 (tried on 13, not useful for resolving the issue)

Ashish Papanai
  • 90
  • 10

1 Answers1

1

The ScatterNDUpdate layer is indeed unsupported for NCS2 since it is not listed in this Supported Layer here.

Your available option is to create a custom layer for VPU that could replace the ScatterNDUpdate functionality. To enable operations not supported by OpenVINO™ out of the box, you need a custom extension for Model Optimizer, a custom nGraph operation set, and a custom kernel for your target device

You may refer to this guide.

Iffa_Intel
  • 106
  • 3