tensorflowqualcommsnpe

ElementWise resolver must implement broadcast method - Qualcomm SNPE

Encountered a problem while trying to run the setup script from Qualcomm SNPE 1.14 to convert a MobileNet_V2 graph into .DLC format.

It passes the optimize_for_inference stage but fails when running convert_to_dlc receiving the following error:

- 123 - ERROR - Conversion failed: ElementWise resolver must implement broadcast method.

I am guessing this is coming from the residual blocks since the add operations are the only elementwise operations found in Mobilenet_V2 expanded_conv blocks.

Solution

  • The error came from using ArgMax, which is not implemented in SNPE. As a solution, cut ArgMax from the graph, convert your model using SNPE and implement ArgMax in the inference testing code.