HCRF_s2_Face_Seg.prototxt

name: "STSR"
layer {
  name: "data"
  type: "HDF5Data"
  top: "data1"
  top: "label"
  hdf5_data_param {
    source: "examples/DecisionTree_CNN/extra/train.txt"
    batch_size: 8
    shuffle: true
  }
  include: { phase: TRAIN }
}

layer {
  name: "data"
  type: "HDF5Data"
  top: "data1"
  top: "label"
  hdf5_data_param {
    source: "examples/DecisionTree_CNN/extra/test.txt"
    batch_size: 8
  }
  include: { phase: TEST }
}


###############################################################
#------------------------ Face segmentation --------------------
########################### Feat Gate #########################

layer {
  name: "Gamma_input"
  bottom: "data1"
  top: "Gamma_input"
  type: "Power"
  power_param {
    power: 0.5
    scale: 1
    shift: 0
  }
}

layer {
  name: "scale_input"
  type: "Scale"
  bottom: "Gamma_input"
  top: "scale_input"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  scale_param {
    filler {
      value: 1
    }
    bias_term: true
    bias_filler {
      value: -0.5
    }
  }
}


######################### s1 Feat bp1 #########################
layer {
  name: "conv1_Feat_bp1"
  type: "Convolution"
  bottom: "scale_input"
  top: "conv1_Feat_bp1"
  param {
    lr_mult: 1
    name: "conv1_bp1_w"
  }
  param {
    lr_mult: 1
    name: "conv1_bp1_b"
  }
  convolution_param {
    num_output: 64
    kernel_size: 3
    stride: 1
    pad: 1
    weight_filler {
      type: "gaussian"
      std: 0.0589
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}

layer {
  name: "relu1_Feat_bp1"
  type: "ReLU"
  bottom: "conv1_Feat_bp1"
  top: "conv1_Feat_bp1"
}

layer {
  name: "conv2_Feat_bp1"
  type: "Convolution"
  bottom: "conv1_Feat_bp1"
  top: "conv2_Feat_bp1"
  param {
    lr_mult: 1
    name: "conv2_bp1_w"
  }
  param {
    lr_mult: 1
    name: "conv2_bp1_b"
  }
  convolution_param {
    num_output: 64
    kernel_size: 3
    stride: 1
    pad: 1
    weight_filler {
      type: "gaussian"
      std: 0.0589
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}

layer {
  name: "relu2_Feat_bp1"
  type: "ReLU"
  bottom: "conv2_Feat_bp1"
  top: "conv2_Feat_bp1"
}

layer {
  name: "pool1"
  type: "Pooling"
  bottom: "conv2_Feat_bp1"
  top: "pool1"
  pooling_param {
    pool: MAX
    kernel_size: 2
    stride: 2
  }
}

######################### s1 Feat bp2 #########################
layer {
  name: "conv1_Feat_bp2"
  type: "Convolution"
  bottom: "pool1"
  top: "conv1_Feat_bp2"
  param {
    lr_mult: 1
    name: "conv1_bp2_w"
  }
  param {
    lr_mult: 1
    name: "conv1_bp2_b"
  }
  convolution_param {
    num_output: 128
    kernel_size: 3
    stride: 1
    pad: 1
    weight_filler {
      type: "gaussian"
      std: 0.0417
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}

layer {
  name: "relu1_Feat_bp2"
  type: "ReLU"
  bottom: "conv1_Feat_bp2"
  top: "conv1_Feat_bp2"
}

layer {
  name: "conv2_Feat_bp2"
  type: "Convolution"
  bottom: "conv1_Feat_bp2"
  top: "conv2_Feat_bp2"
  param {
    lr_mult: 1
    name: "conv2_bp2_w"
  }
  param {
    lr_mult: 1
    name: "conv2_bp2_b"
  }
  convolution_param {
    num_output: 128
    kernel_size: 3
    stride: 1
    pad: 1
    weight_filler {
      type: "gaussian"
      std: 0.0417
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}

layer {
  name: "relu2_Feat_bp2"
  type: "ReLU"
  bottom: "conv2_Feat_bp2"
  top: "conv2_Feat_bp2"
}

layer {
  name: "pool2"
  type: "Pooling"
  bottom: "conv2_Feat_bp2"
  top: "pool2"
  pooling_param {
    pool: MAX
    kernel_size: 2
    stride: 2
  }
}

######################### s1 Feat bp3 #########################
layer {
  name: "conv1_Feat_bp3"
  type: "Convolution"
  bottom: "pool2"
  top: "conv1_Feat_bp3"
  param {
    lr_mult: 1
    name: "conv1_bp3_w"
  }
  param {
    lr_mult: 1
    name: "conv1_bp3_b"
  }
  convolution_param {
    num_output: 256
    kernel_size: 3
    stride: 1
    pad: 1
    weight_filler {
      type: "gaussian"
      std: 0.0295
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}

layer {
  name: "relu1_Feat_bp3"
  type: "ReLU"
  bottom: "conv1_Feat_bp3"
  top: "conv1_Feat_bp3"
}

layer {
  name: "conv2_Feat_bp3"
  type: "Convolution"
  bottom: "conv1_Feat_bp3"
  top: "conv2_Feat_bp3"
  param {
    lr_mult: 1
    name: "conv2_bp3_w"
  }
  param {
    lr_mult: 1
    name: "conv2_bp3_b"
  }
  convolution_param {
    num_output: 256
    kernel_size: 3
    stride: 1
    pad: 1
    weight_filler {
      type: "gaussian"
      std: 0.0295
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}

layer {
  name: "relu2_Feat_bp3"
  type: "ReLU"
  bottom: "conv2_Feat_bp3"
  top: "conv2_Feat_bp3"
}

layer {
  name: "conv3_Feat_bp3"
  type: "Convolution"
  bottom: "conv2_Feat_bp3"
  top: "conv3_Feat_bp3"
  param {
    lr_mult: 1
    name: "conv3_bp3_w"
  }
  param {
    lr_mult: 1
    name: "conv3_bp3_b"
  }
  convolution_param {
    num_output: 256
    kernel_size: 3
    stride: 1
    pad: 1
    weight_filler {
      type: "gaussian"
      std: 0.0295
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}

layer {
  name: "relu3_Feat_bp3"
  type: "ReLU"
  bottom: "conv3_Feat_bp3"
  top: "conv3_Feat_bp3"
}

layer {
  name: "pool3"
  type: "Pooling"
  bottom: "conv3_Feat_bp3"
  top: "pool3"
  pooling_param {
    pool: MAX
    kernel_size: 2
    stride: 2
  }
}

######################### s1 Feat bp4 #########################
layer {
  name: "conv1_Feat_bp4"
  type: "Convolution"
  bottom: "pool3"
  top: "conv1_Feat_bp4"
  param {
    lr_mult: 1
    name: "conv1_bp4_w"
  }
  param {
    lr_mult: 1
    name: "conv1_bp4_b"
  }
  convolution_param {
    num_output: 512
    kernel_size: 3
    stride: 1
    pad: 1
    weight_filler {
      type: "gaussian"
      std: 0.0208
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}

layer {
  name: "relu1_Feat_bp4"
  type: "ReLU"
  bottom: "conv1_Feat_bp4"
  top: "conv1_Feat_bp4"
}

layer {
  name: "conv2_Feat_bp4"
  type: "Convolution"
  bottom: "conv1_Feat_bp4"
  top: "conv2_Feat_bp4"
  param {
    lr_mult: 1
    name: "conv2_bp4_w"
  }
  param {
    lr_mult: 1
    name: "conv2_bp4_b"
  }
  convolution_param {
    num_output: 512
    kernel_size: 3
    stride: 1
    pad: 1
    weight_filler {
      type: "gaussian"
      std: 0.0208
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}

layer {
  name: "relu2_Feat_bp4"
  type: "ReLU"
  bottom: "conv2_Feat_bp4"
  top: "conv2_Feat_bp4"
}

layer {
  name: "conv3_Feat_bp4"
  type: "Convolution"
  bottom: "conv2_Feat_bp4"
  top: "conv3_Feat_bp4"
  param {
    lr_mult: 1
    name: "conv3_bp4_w"
  }
  param {
    lr_mult: 1
    name: "conv3_bp4_b"
  }
  convolution_param {
    num_output: 512
    kernel_size: 3
    stride: 1
    pad: 1
    weight_filler {
      type: "gaussian"
      std: 0.0208
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}

layer {
  name: "relu3_Feat_bp4"
  type: "ReLU"
  bottom: "conv3_Feat_bp4"
  top: "conv3_Feat_bp4"
}

layer {
  name: "pool4"
  type: "Pooling"
  bottom: "conv3_Feat_bp4"
  top: "pool4"
  pooling_param {
    pool: MAX
    kernel_size: 2
    stride: 2
  }
}

######################### Feat_bp5 #########################
layer {
  name: "conv1_Feat_bp5"
  type: "Deconvolution"
  bottom: "pool4"
  top: "conv1_Feat_bp5"
  param {
    lr_mult: 1
    name: "conv1_bp5_w"
  }
  param {
    lr_mult: 1
    name: "conv1_bp5_b"
  }
  convolution_param {
    num_output: 256
    kernel_size: 6
    stride: 2
    pad: 2
    weight_filler {
      type: "gaussian"
      std: 0.0147
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}

layer {
  name: "relu1_Feat_bp5"
  type: "ReLU"
  bottom: "conv1_Feat_bp5"
  top: "conv1_Feat_bp5"
}

layer {
  name: "conv2_Feat_bp5"
  type: "Convolution"
  bottom: "conv1_Feat_bp5"
  top: "conv2_Feat_bp5"
  param {
    lr_mult: 1
    name: "conv2_bp5_w"
  }
  param {
    lr_mult: 1
    name: "conv2_bp5_b"
  }
  convolution_param {
    num_output: 256
    kernel_size: 3
    stride: 1
    pad: 1
    weight_filler {
      type: "gaussian"
      std: 0.0295
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}

layer {
  name: "relu2_Feat_bp5"
  type: "ReLU"
  bottom: "conv2_Feat_bp5"
  top: "conv2_Feat_bp5"
}

layer {
  name: "conv3_Feat_bp5"
  type: "Convolution"
  bottom: "conv2_Feat_bp5"
  top: "conv3_Feat_bp5"
  param {
    lr_mult: 1
    name: "conv3_bp5_w"
  }
  param {
    lr_mult: 1
    name: "conv3_bp5_b"
  }
  convolution_param {
    num_output: 256
    kernel_size: 3
    stride: 1
    pad: 1
    weight_filler {
      type: "gaussian"
      std: 0.0295
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}

layer {
  name: "relu3_Feat_bp5"
  type: "ReLU"
  bottom: "conv3_Feat_bp5"
  top: "conv3_Feat_bp5"
}

######################### Feat_bp3 & Feat_bp5 #########################
layer {
  name: "global_output_Feat_bp3"
  type: "Convolution"
  bottom: "pool3"
  top: "global_output_Feat_bp3"
  param {
    lr_mult: 1
    name: "global_output_bp3_w"
  }
  param {
    lr_mult: 1
    name: "global_output_bp3_b"
  }
  convolution_param {
    num_output: 256
    kernel_size: 1
    stride: 1
    pad: 0
    weight_filler {
      type: "gaussian"
      std: 0.0884
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}

layer {
  name: "concat_Feat_bp3_Feat_bp5"
  type: "Concat"
  bottom: "global_output_Feat_bp3"
  bottom: "conv3_Feat_bp5"
  top: "concat_Feat_bp3_Feat_bp5"
}

######################### Feat_bp6 #########################
layer {
  name: "conv1_Feat_bp6"
  type: "Deconvolution"
  bottom: "concat_Feat_bp3_Feat_bp5"
  top: "conv1_Feat_bp6"
  param {
    lr_mult: 1
    name: "conv1_bp6_w"
  }
  param {
    lr_mult: 1
    name: "conv1_bp6_b"
  }
  convolution_param {
    num_output: 128
    kernel_size: 6
    stride: 2
    pad: 2
    weight_filler {
      type: "gaussian"
      std: 0.0208
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}

layer {
  name: "relu1_Feat_bp6"
  type: "ReLU"
  bottom: "conv1_Feat_bp6"
  top: "conv1_Feat_bp6"
}

layer {
  name: "conv2_Feat_bp6"
  type: "Convolution"
  bottom: "conv1_Feat_bp6"
  top: "conv2_Feat_bp6"
  param {
    lr_mult: 1
    name: "conv2_bp6_w"
  }
  param {
    lr_mult: 1
    name: "conv2_bp6_b"
  }
  convolution_param {
    num_output: 128
    kernel_size: 3
    stride: 1
    pad: 1
    weight_filler {
      type: "gaussian"
      std: 0.0417
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}

layer {
  name: "relu2_Feat_bp6"
  type: "ReLU"
  bottom: "conv2_Feat_bp6"
  top: "conv2_Feat_bp6"
}

######################### Feat_bp2 & SR_s1_bp6 #########################
layer {
  name: "global_output_Feat_bp2"
  type: "Convolution"
  bottom: "pool2"
  top: "global_output_Feat_bp2"
  param {
    lr_mult: 1
    name: "global_output_bp2_w"
  }
  param {
    lr_mult: 1
    name: "global_output_bp2_b"
  }
  convolution_param {
    num_output: 128
    kernel_size: 1
    stride: 1
    pad: 0
    weight_filler {
      type: "gaussian"
      std: 0.125
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}

layer {
  name: "concat_Feat_bp2_Feat_bp6"
  type: "Concat"
  bottom: "global_output_Feat_bp2"
  bottom: "conv2_Feat_bp6"
  top: "concat_Feat_bp2_Feat_bp6"
}

######################### Feat_bp7 #########################
layer {
  name: "conv1_Feat_bp7"
  type: "Deconvolution"
  bottom: "concat_Feat_bp2_Feat_bp6"
  top: "conv1_Feat_bp7"
  param {
    lr_mult: 1
    name: "conv1_bp7_w"
  }
  param {
    lr_mult: 1
    name: "conv1_bp7_b"
  }
  convolution_param {
    num_output: 64
    kernel_size: 6
    stride: 2
    pad: 2
    weight_filler {
      type: "gaussian"
      std: 0.0295
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}

layer {
  name: "relu1_Feat_bp7"
  type: "ReLU"
  bottom: "conv1_Feat_bp7"
  top: "conv1_Feat_bp7"
}

layer {
  name: "conv2_Feat_bp7"
  type: "Convolution"
  bottom: "conv1_Feat_bp7"
  top: "conv2_Feat_bp7"
  param {
    lr_mult: 1
    name: "conv2_bp7_w"
  }
  param {
    lr_mult: 1
    name: "conv2_bp7_b"
  }
  convolution_param {
    num_output: 64
    kernel_size: 3
    stride: 1
    pad: 1
    weight_filler {
      type: "gaussian"
      std: 0.0589
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}

layer {
  name: "relu2_Feat_bp7"
  type: "ReLU"
  bottom: "conv2_Feat_bp7"
  top: "conv2_Feat_bp7"
}

######################### Feat_bp1 & SR_s1_bp7 #########################
layer {
  name: "global_output_Feat_bp1"
  type: "Convolution"
  bottom: "pool1"
  top: "global_output_Feat_bp1"
  param {
    lr_mult: 1
    name: "global_output_bp1_w"
  }
  param {
    lr_mult: 1
    name: "global_output_bp1_b"
  }
  convolution_param {
    num_output: 64
    kernel_size: 1
    stride: 1
    pad: 0
    weight_filler {
      type: "gaussian"
      std: 0.1768
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}

layer {
  name: "concat_Feat_bp1_Feat_bp7"
  type: "Concat"
  bottom: "global_output_Feat_bp1"
  bottom: "conv2_Feat_bp7"
  top: "concat_Feat_bp1_Feat_bp7"
}

######################### Feat_bp8 #########################
layer {
  name: "conv1_Feat_bp8"
  type: "Deconvolution"
  bottom: "concat_Feat_bp1_Feat_bp7"
  top: "conv1_Feat_bp8"
  param {
    lr_mult: 1
    name: "conv1_bp8_w"
  }
  param {
    lr_mult: 1
    name: "conv1_bp8_b"
  }
  convolution_param {
    num_output: 32
    kernel_size: 6
    stride: 2
    pad: 2
    weight_filler {
      type: "gaussian"
      std: 0.0417
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}

layer {
  name: "relu1_Feat_bp8"
  type: "ReLU"
  bottom: "conv1_Feat_bp8"
  top: "conv1_Feat_bp8"
}

layer {
  name: "conv2_Feat_bp8"
  type: "Convolution"
  bottom: "conv1_Feat_bp8"
  top: "conv2_Feat_bp8"
  param {
    lr_mult: 1
    name: "conv2_bp8_w"
  }
  param {
    lr_mult: 1
    name: "conv2_bp8_b"
  }
  convolution_param {
    num_output: 32
    kernel_size: 3
    stride: 1
    pad: 1
    weight_filler {
      type: "gaussian"
      std: 0.0833
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}

layer {
  name: "relu2_Feat_bp8"
  type: "ReLU"
  bottom: "conv2_Feat_bp8"
  top: "conv2_Feat_bp8"
}

######################### Feat splice #########################

layer {
  name: "conv_SR_feat_map"
  type: "Convolution"
  bottom: "conv2_Feat_bp8"
  top: "conv_SR_feat_map"
  param {
    lr_mult: 1
    name: "conv_SR_feat_map_w"
  }
  param {
    lr_mult: 1
    name: "conv_SR_feat_map_b"
  }
  convolution_param {
    num_output: 32
    kernel_size: 3
    stride: 1
    pad: 1
    weight_filler {
      type: "gaussian"
      std: 0.0833
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}

layer {
  name: "relu_conv_SR_feat_map"
  type: "ReLU"
  bottom: "conv_SR_feat_map"
  top: "conv_SR_feat_map"
}

layer {
  name: "conv_SR_feat_bp0"
  type: "Convolution"
  bottom: "conv2_Feat_bp1"
  top: "conv_SR_feat_bp0"
  param {
    lr_mult: 1
    name: "conv_SR_feat_bp0_w"
  }
  param {
    lr_mult: 1
    name: "conv_SR_feat_bp0_b"
  }
  convolution_param {
    num_output: 32
    kernel_size: 3
    stride: 1
    pad: 1
    weight_filler {
      type: "gaussian"
      std: 0.0833
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}

layer {
  name: "SR_feat_fine"
  type: "Eltwise"
  bottom: "conv_SR_feat_map"
  bottom: "conv_SR_feat_bp0"
  top: "SR_feat_fine"
  eltwise_param {
      operation: SUM
  }
}

######################### Feat concatenate #########################
layer {
  name: "Feat"
  type: "Convolution"
  bottom: "SR_feat_fine"
  top: "Feat"
  param {
    lr_mult: 1
    name: "Feat_w"
  }
  param {
    lr_mult: 1
    name: "Feat_b"
  }
  convolution_param {
    num_output: 6
    kernel_size: 3
    stride: 1
    pad: 1
    weight_filler {
      type: "gaussian"
      std: 0.1925
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}

######################### Feat LOSS stage 1 #########################
layer {
  name: "GT_Feat_loss"
  type: "SoftmaxWithLoss"
  bottom: "Feat"
  bottom: "label"
  top: "GT_Feat_loss"
  loss_weight: 1
  loss_param {
    #ignore_label: 0
    normalize: false
  }
}