frame

Howdy, Stranger!

It looks like you're new here. If you want to get involved, click one of these buttons!

Sign In

Howdy, Stranger!

It looks like you're new here. If you want to get involved, click one of these buttons!

Error happened when I try to use mobilenet v2 in movidius stick.

I am so sorry to post so many "Look for help" on the same day,
But I am really trouble in some problems for a long time when I try to put a mobiltnet v2 in Movidius stick,

the follow is some information

SDK version : 2.05
NET : mobilenet v2 (https://github.com/shicai/MobileNet-Caffe)


step : 0
ok, then I use mvNCCompile to translate mobilenet v2 to graph,
it was successfully, I get the graph file.


step : 1
then , I use mvNCCheck to make sure mobilenet v2 have be converted normally,
I get nan result,
it seems there some problems there,
TO find what has happened,
I use mvNCCheck with input node = input, output node = each layer,
and I find after the data pass a bn layer, data = nan,


I guess there may be an overflow happened,
So I go to the https://developer.movidius.com/docs to find more information,
then I get this,

_Input/output data types
In NCAPI v1, input data for inferences had to be 16 bit floating point data type.

In NCAPI v2, input data for inferences can be 16 bit or 32 bit floating point data type. The default data type is FP32. The data type is set for each Fifo using Fifo.set_option() with FifoDataType.

Note: If the input FifoDataType is configured to FP32, the API will convert the data to the FP16 data type automatically when performing inferences. If the output FifoDataType is configured to 32FP, the API will convert the output back to the FP32 data type._

OK , it tells me NCAPI V2 would convert FP32 to FP16 to use , it means NCAPI v2 could deal with 32 bit data,the same as a computer,
it means there shouldn't have any overflow in the stick
,
so,
I wonder whether mvNCCheck tool can deal with 32 bit, or maybe mvNCCheck only support 16 bit which leads an overflow

thaks again for your finishing reading my problem
Really, thanks.

Sign In or Register to comment.