Train

#This network is like a pre-alpha version. I just sort of wrote it out of the blue.
#data size is 227x227x3
layer {
  name: "data"
  type: "Data"
  top: "data"
  include {
    phase: TRAIN
  }
  transform_param {
    scale: 1
  }
  data_param {
    source: "C:/Users/apple/models/wilde"
    batch_size: 1 #about 20 iterations per second on a gtx1050
    backend: LMDB
  }
}
layer {
  name: "conv1_a"
  type: "Convolution"
  bottom: "data"
  top: "conv1_a"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 16
    pad: 2
    kernel_size: 5
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  name: "relu1_a"
  type: "ELU"
  bottom: "conv1_a"
  top: "conv1_a"
}
layer {
  name: "conv2_a"
  type: "Convolution"
  bottom: "conv1_a"
  top: "conv2_a"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 16
    pad: 2
    kernel_size: 5
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  name: "relu2_a"
  type: "ELU"
  bottom: "conv2_a"
  top: "conv2_a"
}
layer {
  name: "conv3_a"
  type: "Convolution"
  bottom: "conv2_a"
  top: "conv3_a"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 16
    pad: 1
    kernel_size: 3
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  name: "relu3_a"
  type: "ELU"
  bottom: "conv3_a"
  top: "conv3_a"
}
layer {
  name: "pool3_a"
  type: "Pooling"
  bottom: "conv3_a"
  top: "pool3_a"
  pooling_param {
    pool: MAX
    kernel_size: 3
    stride: 2
  }
}
layer {
  name: "conv1_b"
  type: "Convolution"
  bottom: "pool3_a"
  top: "conv1_b"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 32
    pad: 2
    kernel_size: 5
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  name: "relu1_b"
  type: "ELU"
  bottom: "conv1_b"
  top: "conv1_b"
}
layer {
  name: "conv2_b"
  type: "Convolution"
  bottom: "conv1_b"
  top: "conv2_b"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 32
    pad: 2
    kernel_size: 5
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  name: "relu2_b"
  type: "ELU"
  bottom: "conv2_b"
  top: "conv2_b"
}
layer {
  name: "conv3_b"
  type: "Convolution"
  bottom: "conv2_b"
  top: "conv3_b"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 32
    pad: 1
    kernel_size: 3
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  name: "relu3_b"
  type: "ELU"
  bottom: "conv3_b"
  top: "conv3_b"
}
layer {
  name: "pool3_b"
  type: "Pooling"
  bottom: "conv3_b"
  top: "pool3_b"
  pooling_param {
    pool: MAX
    kernel_size: 3
    stride: 2
  }
}
layer {
  name: "conv1_c"
  type: "Convolution"
  bottom: "pool3_b"
  top: "conv1_c"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 64
    pad: 2
    kernel_size: 5
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  name: "relu1_c"
  type: "ELU"
  bottom: "conv1_c"
  top: "conv1_c"
}
layer {
  name: "conv2_c"
  type: "Convolution"
  bottom: "conv1_c"
  top: "conv2_c"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 64
    pad: 2
    kernel_size: 5
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  name: "relu2_c"
  type: "ELU"
  bottom: "conv2_c"
  top: "conv2_c"
}
layer {
  name: "conv3_c"
  type: "Convolution"
  bottom: "conv2_c"
  top: "conv3_c"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 64
    pad: 1
    kernel_size: 3
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  name: "relu3_c"
  type: "ELU"
  bottom: "conv3_c"
  top: "conv3_c"
}
layer {
  name: "pool3_c"
  type: "Pooling"
  bottom: "conv3_c"
  top: "pool3_c"
  pooling_param {
    pool: MAX
    kernel_size: 3
    stride: 2
  }
}
layer {
  name: "conv1_d"
  type: "Convolution"
  bottom: "pool3_c"
  top: "conv1_d"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 96
    pad: 2
    kernel_size: 5
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  name: "relu1_d"
  type: "ELU"
  bottom: "conv1_d"
  top: "conv1_d"
}
layer {
  name: "conv2_d"
  type: "Convolution"
  bottom: "conv1_d"
  top: "conv2_d"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 96
    pad: 2
    kernel_size: 5
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  name: "relu2_d"
  type: "ELU"
  bottom: "conv2_d"
  top: "conv2_d"
}
layer {
  name: "conv3_d"
  type: "Convolution"
  bottom: "conv2_d"
  top: "conv3_d"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 96
    pad: 1
    kernel_size: 3
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  name: "relu3_d"
  type: "ELU"
  bottom: "conv3_d"
  top: "conv3_d"
}
layer {
  name: "pool3_d"
  type: "Pooling"
  bottom: "conv3_d"
  top: "pool3_d"
  pooling_param {
    pool: MAX
    kernel_size: 3
    stride: 2
  }
}
layer {
  name: "conv1_e"
  type: "Convolution"
  bottom: "pool3_d"
  top: "conv1_e"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 64
    pad: 2
    kernel_size: 5
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  name: "relu1_e"
  type: "ELU"
  bottom: "conv1_e"
  top: "conv1_e"
}
layer {
  name: "conv2_e"
  type: "Convolution"
  bottom: "conv1_e"
  top: "conv2_e"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 64
    pad: 2
    kernel_size: 5
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  name: "relu2_e"
  type: "ELU"
  bottom: "conv2_e"
  top: "conv2_e"
}
layer {
  name: "conv3_e"
  type: "Convolution"
  bottom: "conv2_e"
  top: "conv3_e"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 64
    pad: 1
    kernel_size: 3
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  name: "relu3_e"
  type: "ELU"
  bottom: "conv3_e"
  top: "conv3_e"
}
layer {
  name: "deconv_fc1"
  type: "InnerProduct"
  bottom: "conv3_e"
  top: "deconv_fc1"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  inner_product_param {
    num_output: 512
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "relu_deconv_fc1"
  type: "ELU"
  bottom: "deconv_fc1"
  top: "deconv_fc1"
}
layer {
  name: "deconv_fc2"
  type: "InnerProduct"
  bottom: "deconv_fc1"
  top: "deconv_fc2"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  inner_product_param {
    num_output: 1024
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "relu_deconv_fc2"
  type: "ELU"
  bottom: "deconv_fc2"
  top: "deconv_fc2"
}
layer {
  name: "deconv_fc3_"
  type: "InnerProduct"
  bottom: "deconv_fc2"
  top: "deconv_fc3_"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  inner_product_param {
    num_output: 12544
# this is probably a bit excessive but it can be reshaped back into the size of layer "conv3_e" which is 1x64x14x14 which aids a lot in creating a spicy decoder.
#Other values could work, as long as it can be reshaped back into Cx14x14
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "relu_deconv_fc3_"
  type: "ELU"
  bottom: "deconv_fc3_"
  top: "deconv_fc3_"
}
layer {
  name: "reshape_fc"
  type: "Reshape"
  bottom: "deconv_fc3_"
  top: "reshape_fc"
  reshape_param {
    shape {
      dim: 1
      dim: -1
      dim: 14
      dim: 14
    }
  }
}
layer {
  name: "conv1_e_t"
  type: "Deconvolution"
  bottom: "reshape_fc"
  top: "conv1_e_t"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  convolution_param {
    num_output: 96
    pad: 1
    kernel_size: 3
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  name: "relu1_e_t"
  type: "ELU"
  bottom: "conv1_e_t"
  top: "conv1_e_t"
}
layer {
  name: "pool3_d_t"
  type: "Resize"
  bottom: "conv1_e_t"
  top: "pool3_d_t"
  resize_param {
    interpolation: LINEAR
    height: 28
    width: 28
  }
}
layer {
  name: "conv1_d_t"
  type: "Deconvolution"
  bottom: "pool3_d_t"
  top: "conv1_d_t"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  convolution_param {
    num_output: 64
    pad: 1
    kernel_size: 3
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  name: "relu1_d_t"
  type: "ELU"
  bottom: "conv1_d_t"
  top: "conv1_d_t"
}
layer {
  name: "pool3_c_t"
  type: "Resize"
  bottom: "conv1_d_t"
  top: "pool3_c_t"
  resize_param {
    interpolation: LINEAR
    height: 56
    width: 56
  }
}
layer {
  name: "conv1_c_t"
  type: "Deconvolution"
  bottom: "pool3_c_t"
  top: "conv1_c_t"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  convolution_param {
    num_output: 32
    pad: 1
    kernel_size: 3
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  name: "relu1_c_t"
  type: "ELU"
  bottom: "conv1_c_t"
  top: "conv1_c_t"
}
layer {
  name: "pool3_b_t"
  type: "Resize"
  bottom: "conv1_c_t"
  top: "pool3_b_t"
  resize_param {
    interpolation: LINEAR
    height: 113
    width: 113
  }
}
layer {
  name: "conv1_b_t"
  type: "Deconvolution"
  bottom: "pool3_b_t"
  top: "conv1_b_t"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  convolution_param {
    num_output: 16
    pad: 1
    kernel_size: 3
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  name: "relu1_b_t"
  type: "ELU"
  bottom: "conv1_b_t"
  top: "conv1_b_t"
}
layer {
  name: "pool3_a_t"
  type: "Resize"
  bottom: "conv1_b_t"
  top: "pool3_a_t"
  resize_param {
    interpolation: LINEAR
    height: 227
    width: 227
  }
}
layer {
  name: "conv1_a_t"
  type: "Deconvolution"
  bottom: "pool3_a_t"
  top: "conv1_a_t"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  convolution_param {
    num_output: 3 #Change to 4 if you're feeling ambitious and want to train on images with an alpha (transparency channel(.
    pad: 1
    kernel_size: 3
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  name: "relu1_a_t"
  type: "ELU"
  bottom: "conv1_a_t"
  top: "conv1_a_t"
}
layer {
  name: "mse_loss"
  type: "EuclideanLoss"
  bottom: "data"
  bottom: "conv1_a_t"
  top: "mse_loss"
  loss_weight: 1e-09 #scales down the loss into something more acceptable; it probably comes from the fact that either the "data" or the "conv1_at" need more than just an "ELU" layer for normalization.
}