Run parts of network on different devices

[Broderick__Colin]

I'm so close to getting my model running on the NCS2! Unfortunately I can't make the final leap because one of my layers is not supported. I'd appreciate it if someone could answer one or both of these questions:

1. Is it possible to execute some layers on the CPU, and some layers on the MYRIAD chip? It's only one layer so I'm not worried about the performance hit.

2. The unsupported layer is the input layer, the XML of which is:

<layer id="0" name="input" precision="FP32" type="Input">
    <output>
        <port id="0">
            <dim>512</dim>
            <dim>6</dim>
            <dim>200</dim>
        </port>
    </output>
</layer>

I realize that FP32 is not supported. I will rebuild it in FP16. I don't think that's the reason for the issue though because all the other layers are current FP32 as well, and they do not come up as unsupported.

I'm working in Python by the way.

Thanks.

Edit: To clear, it runs fine on the CPU. And I have since rebuilt it in FP16 but the problem persists.

[SIRIGIRI_V_Intel]

Hi Colin,

You can try using -d MULTI:MYRIAD,CPU for running the model on two devices.

Also, You can try to implement the unsupported layers as custom layer.

Regards,

Ram prasad

[Broderick__Colin1]

Thank you, I have solved the problem.

I think I was misunderstanding the need to check for unsupported layers. I was doing so for the MYRIAD and it was always telling me the first layer, whatever it was, was unsupported. I've just stopped doing that check and everything is fine.