possible mismatch within tensorDesc for hello_classification c sample

[guoyejun]

Hi,

i don't quite understand the following code, is there a mismatch? thanks.

https://github.com/openvinotoolkit/openvino/blob/2020/inference-engine/ie_bridges/c/samples/hello_classification/main.c#L140

    dimensions_t dimens = {4, {1, (size_t)img.mat_channels, (size_t)img.mat_height, (size_t)img.mat_width}};
    tensor_desc_t tensorDesc = {NHWC, dimens, U8};

tensorDesc says that NHWC is used, but dimens contains NCHW (n=1, c=img.mat_channels, h=img.mat_height, w=img.mat_width).

Did I miss anything? thanks.

[ArtemSkrebkov]

Hi @guoyejun
The order of dims in the IE library is fixed and it is always NCHW for 4-D data. If you know that your data has NHWC layout you need to specify the layout field of TensorDesc but no need to change the order of the dims. It is, again, NCHW.

Have tried to find this info in the documentation but couldn't :(
You can rely on implementation of getLayoutByDims. It returns an order of channels based on a vector of dimensions passed.

Hope my comment makes things a bit clearer.

[guoyejun]

thanks,

will openvino add an internal convert from NHWC to NCHW if we set NHWC layout? thanks.

[ArtemSkrebkov]

Yes. If a network expects to get NCHW but a user provides NHWC. It is a responsibility of a concrete plugin to do a conversion.

[guoyejun]

thanks, and just curious why NHWC needs to be set twice, once here, and another place at https://github.com/openvinotoolkit/openvino/blob/2020/inference-engine/ie_bridges/c/samples/hello_classification/main.c#L109

[ilya-lavrenov]

@guoyejun it's not set twice. First in line 109 we set layout which plugin will except (and Core::LoadNetwork will compile for such layout), in line 140 we create a blob with such layout.

[guoyejun]

thanks