Deconvolution 与convolution

版权声明：本文为博主原创文章，如需转载请附上博文链接 https://blog.csdn.net/wendygelin/article/details/88549795

Deconvolution和convolution相反，有upsample的作用。如下面代码中，   

    pad: 1
    kernel_size: 4
    stride: 2

达到的效果是，将图的长和宽都增大了一倍。

layer {
  name: "deconv5"
  type: "Deconvolution"
  bottom: "conv6_1"
  top: "deconv5"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  convolution_param {
    num_output: 192
    pad: 1
    kernel_size: 4
    stride: 2
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
    }
    engine: CUDNN
  }
}

而convolution，可以将图缩小，如下代码，

    pad: 1
    kernel_size: 3
    stride: 2

则图的长和宽都缩小了一半。

layer {
  name: "conv6"
  type: "Convolution"
  bottom: "conv5_1"
  top: "conv6"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  convolution_param {
    num_output: 384
    pad: 1
    kernel_size: 3
    stride: 2
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
    }
    engine: CUDNN
  }
}