Xu Hai

我跟你遇到了一样的问题，怎么解决的？

i enable the both CAFFE2_MOBILE and CAFFE2_USE_MPSCNN. then compile with a error: caffe2/caffe2/mobile/contrib/ios/mpscnn/mpscnn.mm:24:10: fatal error: 'caffe2/mobile/fb/aloha/headers.h' file not found #include "caffe2/mobile/fb/aloha/headers.h"

the libcaffe2.a was succussfully built. But when I use the lib, I got a error: ``` Undefined symbols for architecture arm64: "_OBJC_CLASS_$_MPSCNNSoftMax", referenced from: objc-class-ref in libcaffe2.a(mpscnn.mm.o) "_OBJC_CLASS_$_MPSCNNPoolingMax", referenced from:...

@landiaokafeiyan Does your training prototxt work? How about the solver.txt?

after reading the paper, I found the loss function is two softmax loss function addup together with the weight of 0.4. Take pspnet101_VOC2012_473.prototxt as an example, I guess the loss...

Here is the training proto, I can train by using these protos. https://github.com/SoonminHwang/caffe-segmentation/tree/master/pspnet/models

you need an initial model, just fine-tune and you will get right result, do not train from scratch.

@ThienAnh You can use pspnet101_VOC2012.caffemodel for tuning or training. Why use initial model? You can not train a model without init model, otherwise it will not converge.

Is this one? PSPNet/evaluation/prototxt/pspnet101_VOC2012_473.prototxt change the data layer to the training data?