All backward will be zeros if "lr_mult" of Deconvolution layer is assigned as non-zero. about crfasrnn

[bittnt]

I do not think it matters with "lr_mult".

As far as I know, for new prototxt, it might worthwhile to check if the weights for deconvolution layers are initialized properly. If you change the name of deconvolution layer, the weights might just initialize to zero. In new code base, you can set the weight initialization for deconvolution layer, but it does not exist in the previous code base.

[thesby]

@bittnt Thank you for your answer. I just want to train a new model without any restoring, and all the convolutional / deconvolutional parameters are initialized with xavier method. If "lr_mult" is non-zero, all the backward blobs become zeros. It seems strange.

[thesby]

@bittnt I am sorry that I have merged your caffe version to the latest caffe before your granting and published CRFasRNN-merge. If you are not satisfy with this, I will remove the repository as soon as possible.

[thesby]

I rebuild the caffe, and it works right now. I think there is some other error, not lr_mult.

[bittnt]

deconvolutional parameters should be initialized with bilinear weights. e.g.

layer {
  name: "upsample_example"
  type: "Deconvolution"
  bottom: "upscore2"
  top: "upscore4"
  param {
    lr_mult: 0
  }
        convolution_param{
                bias_term: false
                num_output: 21
                kernel_size: 4
                stride: 2
                pad: 1
                weight_filler: { type: "bilinear" }
        }
}

[bittnt]

It is great to hear that you successfully merged the code with caffe upstream.

I have also done a version in crfrnn branch of our caffe fork.

[thesby]

Wow, bilinear should be better than xavier for expanding. I will try this. Thank you