FCN code giving odd results

Hi,

I am training FCN using my data for segmentation. I am facing a problem. When I use “fcnLayers” to train FCN on my data, it gives me proper segmented output.

But when I open the network (FCN) in deep network designer and generate the code of FCN from there and use that code to train FCN on my data it gives me very odd result. Can anyone guide me what is the reason of that?

This is segmented output using "fcnLayers" command.

The sgmented output using fcn code ( genertes from the deep network designer)

Answer 1

[Birju Patel]

The fcnLayers functions returns a network with image net trained weights. When you generate code from Deep Network Designer, make sure to select “with initial parameters”. This will store the pretrained weights of the FCN network created by the app. If you don't do that, then the weights are left randomly initialized when you train. That is probably why you're getting different results.

[Shoaib Ali]

Thanks a alot.