Get wrong result after converting the model to coreML

[huangxf14]

Hello! I have used the caffe model you provided to get a semantic segmentation(two classes) model (just change to model into a FCN model). Then I convert the caffe model to coreML model. I can get reasonable result on iphone X but the result is not so good that is different from the result I get from caffe model on Ubuntu. Have you met problems like this? Is there any difference between caffe and coreML that matters? I am confused with it and thanks for you reply.

[huangxf14]

And I meet the same problem when I run the coreML model just on simulator. To figure out whether the 'group' in convolution matters , I detele all the group in the caffe prototxt and fill the caffemodel with 0. It shows that the model without group will give the same result.

[hollance]

This happens often enough that I wrote a blog post about it: http://machinethink.net/blog/help-core-ml-gives-wrong-output/

[huangxf14]

Thanks for your reply. I have fixed the bug about input. But I still find a little difference between the result I get from coreML and caffe. The two images below are the result I get from caffe and coreML. I use the model to segment human from the image. As you can see, there is a little difference. Is it because the different implementation between caffe and coreML?

[hollance]

This might be caused by the fact that the GPU in iOS devices using 16-bit floating point instead of the 32-bit floating point used by Caffe. I would try running the Core ML model in CPU mode (see the MLPredictionOptions) and comparing the output then.

[huangxf14]

I have try to run the model on linux server, simulator, iphone X. And I get three different results. Will it run in CPU mode in the simulator? (The two images I show are from linux server and simulator)

[hollance]

Yes, simulator uses the CPU not the GPU.