Hello, I have a question about the structure. In your paper, there is no other layers bwtween the two conv layers with shared weights , but as we know two conv layers with no act layer in between equal to one conv layer, which means you can just use one conv layer of shared weights.

Thanks for your sharing!
When I run mx.model.load_checkpoint('CRU-Net-116-ATx28x14-32x4d',0),There are some error:

~/anaconda3/envs/mxnet3/lib/python3.6/site-packages/mxnet/symbol/symbol.py in load(fname)
2419 raise TypeError('fname need to be string')
2420 handle = SymbolHandle()
-> 2421 check_call(_LIB.MXSymbolCreateFromFile(c_str(fname), ctypes.byref(handle)))
2422 return Symbol(handle)
2423

~/anaconda3/envs/mxnet3/lib/python3.6/site-packages/mxnet/base.py in check_call(ret)
143 """
144 if ret != 0:
--> 145 raise MXNetError(py_str(_LIB.MXGetLastError()))
146
147 if sys.version_info[0] < 3:

MXNetError: Failed loading Op conv4_x__1_c3x3-b(1)__conv of type ChannelwiseConvolution: [09:17:56] src/core/op.cc:55: Check failed: op != nullptr Operator ChannelwiseConvolution is not registered

I hope your help,Thanks.

If using the architecture of CRU-Net-56@x14 (32*4d) without sharing parameters, how about the performance?
Is it higher than 21.9 (top-1 error)? this comparison seems not listed in the paper.

Have you experimented CRU approach on semantic segmentation tasks like https://github.com/itijyou/ademxapp?

I use your ChannelwiseConvolution to implement mobilenet. However i only can get (2s/image without mkl, 0.90 with mkl) in cpu while tensorflow-mobilenet is 0.059s/image. So i ask for some idea to improve speed in cpu.

There are links only to Trained model. I could not find the code for training or defining the CRU model. Please let me know where can I find the source code.