OpenVINO model optimizer incorrectly convert MxNet's broadcast_mul

[Xu_W_]

Hi,

I'm converting a MxNet model into IR format using mo_mxnet.py script.

The script converts MxNet's "broadcast_mul" operators into "EltWise" layers but it doesn't set the operation=mul property. So after convertion, the element-wise multiply operation becomes the element-wise add operation, and produce wrong results. If I manually add operation=mul property in the result XML file, the converted model can produce correct results.

It seems there is a bug in optimizer script, hope it will get fixed.

Thanks.

[Monique_J_Intel]

Hi Xu,

 Can you give me the model or if it's a standard public model tell me the name and the command used to convert the model via Model Optimizer to reproduce the issue?

Kind Regards,

Monique Jones

[Xu_W_]

Hi Jones,

You may download the model here:

https://pan.baidu.com/s/1-uHh7WGun3-NESFWDMUg5Q

The command used to convert the model is:

mo_mxnet.py --input_symbol=areca_full-symbol.json --input_shape=(1,3,1152,1536) --input_model=areca_full-0000.params

Thanks.