Gxllii / U-NET-for-Caffe

U-Net with upsampling layer under caffe

U-NET-for-Caffe

U-Net with upsampling layer under caffe

1. copy the src/caffe/layers/upsample_layer.cpp and src/caffe/layers/upsampling_layer.cu and include/caffe/layers/upsample_layer.h to your caffe

2. edit the src/caffe/proto/caffe.proto.

   add a declaration of parameters used in upsample_layer and give the parameter a unique ID.

   add a line inside message LayerParameter {} .

   // Parameter for upsample_layer. optional UpsampleParameter upsample_param = 152;

   add a line after message LayerParameter {}.

   // message for upsample parameter

   message UpsampleParameter { optional int32 scale = 1 [default = 1]; }

3. re-build the source code of caffe in vs.

   if you use python interference, don't forget copy the released file in \Build\x64\Release\pycaffe\caffe to site-packages of python path.

4. I modified the prototxt file of unet network with deconvolution layer, because its params are difficult to initialize. I always have trouble in training deconvolution layer.

   the deconvolution layer is replaced by upsampling layer + convolution layer with 2*2 kernel, which is as same structure as unet in keras.

5. you should modify the parameter offset in crop layer to make sure the bottom0 blob is cropped in center. the offsite=(bottom0-bottom1)/2, note that bottom0 is the one to be cropped and it is larger.

6. you can try your own data now!