weight initialization

[evagian]

Hello :)
Could you please help me with something?...
It is not clear to me in which part of your code you initialize the weights of the CNN.

[anlijuncn]

It's also confused to me too. I recommend you to read https://github.com/SaoYan/DnCNN-PyTorch
I don't know if we need some skills to initialize these weights with BN layers. If you have reached to the final result in paper. Please SHARE on you github!! I am now working on it, but only could get PSNR 29.14 for DnCNN-S. :)

[evagian]

Thank you @anlijuncn

I think that they are using the default tf.layers.conv2d initializer which is Xavier but I am not sure about that.

[anlijuncn]

Yes, in crisb-dut's code he is using default initializer. I don't know whether methods could improve the result. The author ZhangKai has released Pytorch and Keras version code on github/cszn.

[eastchun]

Default kernel initializer for tf.layer.conv2d is "glorot_uniform_initializer".

Please refer to code lines '315 - 317' in : https://github.com/tensorflow/tensorflow/blob/master/tensorflow/python/ops/variable_scope.py