FP16 with NCS

[sui_g_]

Hi,

I use --data_type=fp32 to convert a model, input precision and output precision both using FP32. It is working fine.

However, if I use --data_type=fp16 to convert model to use in NCS, (keeping input and output precision same FP32,) I got invalid results.

Does it mean that I have convert my input and output precision to FP16 as well? Is there any example to do so? I couldn't find any examples to illustrate this type of conversion (FP16 -> FP32, FP32->FP16).

I am quite sure my inputs and model weights are within the range of FP16. Mostly 0 ~ 1.

Thanks for help~

[Severine_H_Intel]

Hi Sui, 

the input does not need to be converted.  I think it depends on the topology and if it is supported and/or validated on NCS. Which topology do you have?

Best, 

Severine

[sui_g_]

Hi, Severine:

Thanks a lot for your quick helps~ That's is what i think too,  it may be  the topology. Because I also try directly using mvNCCheck for the caffe model and it gives error. It seems that NCS does not support caffe with multiple concat layers chaining together. In my model, the intermediate concat layer output is also used by other path too.

Thanks,

Sui Guan