3D conv output is 0

[Peter Neher]

Hi everyone,

I am trying to perform segmentation using a 3D U-Net. 2D worked fine, but now I am trying to segment on 3D input patches. Unfortunately, already the output of the first convolutional layer is always zero. I am training using the python interface:

solver = caffe.get_solver_from_file("solver_file.prototxt")

for i in range(iterations) :
  train_data = get_batch()
  solver.net.blobs['data'].data[...] = train_data['data']
  solver.net.blobs['target'].data[...] = train_data['seg']
  solver.step(1)

When I collect the output of the first conv layer (solver.net.blobs['contr_1_1'].data) it is always 0. I am using the opencl branch of caffe since this apparently supports 3D for all layers.

I attached the solver and network files. Any suggestions what might go wrong here?

Cheers,
Peter