在Caffe上進行遷移學習

在 Caffe-SSD 上進行遷移學習

目錄

• 在 Caffe-SSD 上進行遷移學習
  • [TOC]
  • 機器環境
  • 准備數據集
    • 得到 trainval.txt
    • 得到 lmdb 文件
  • 准備訓練
  • 進行訓練
  • 其他

機器環境

OS: Ubuntu16.04
GPU: TITAN Xp

編譯Caffe-SSD時，我們沒有選擇 python 接口。不過事后來看也許添加 python 接口會更方便。

准備數據集

得到 trainval.txt

由於習慣，我們將圖片和標注都按照VOC數據集的格式組織，放到caffe-ssd/data/VOC0712目錄下。

+ caffe-ssd:
    + data:
        ...
        + VOC0712:
            - create_list.sh
            - create_data.sh
            - labelmap_emotor.prototxt
            + VOCdevkit:
                + Anotations:
                    - all .xml
                + ImageSets:
                    + Main:
                        - emotor_train.txt
                        - emotor_val.txt
                + JPEGImages:
                    - all pics

首先執行create_list.sh創建trainval.txt。需要根據自己文件夾的結構修改create_list.sh中的路徑。
trainval.txt中的內容如下（一行例子）：

VOC2012/JPEGImages/486.jpg VOC2012/Annotations/486.xml

作用是記錄.jpg和對應的.xml的位置。

我沒有創建用於測試的數據，測試數據的生成方法如出一轍。

得到 lmdb 文件

caffe提供了用於轉換帶標注的數據集的工具。位於/caffe-ssd/build/tools下的conver_annoset。實際的位置跟編譯caffe時候的設置有關。

convert_annoset: Convert a set of images and annotations to the leveldb/lmdb format used as input for Caffe.
Usage:
    convert_annoset [FLAGS] ROOTFOLDER/ LISTFILE DB_NAME


  Flags from tools/convert_annoset.cpp:
    -anno_type (The type of annotation {classification, detection}.)
      type: string default: "classification"
    -backend (The backend {lmdb, leveldb} for storing the result) type: string
      default: "lmdb"
    -check_label (When this option is on, check that there is no duplicated
      name/label.) type: bool default: false
    -check_size (When this option is on, check that all the datum have the same
      size) type: bool default: false
    -encode_type (Optional: What type should we encode the image as
      ('png','jpg',...).) type: string default: ""
    -encoded (When this option is on, the encoded image will be save in datum)
      type: bool default: false
    -gray (When this option is on, treat images as grayscale ones) type: bool
      default: false
    -label_map_file (A file with LabelMap protobuf message.) type: string
      default: ""
    -label_type (The type of annotation file format.) type: string
      default: "xml"
    -max_dim (Maximum dimension images are resized to (keep same aspect ratio))
      type: int32 default: 0
    -min_dim (Minimum dimension images are resized to (keep same aspect ratio))
      type: int32 default: 0
    -resize_height (Height images are resized to) type: int32 default: 0
    -resize_width (Width images are resized to) type: int32 default: 0
    -shuffle (Randomly shuffle the order of images and their labels) type: bool
      default: false

在我們的例子中參數設置如下：

./convert_annoset --anno_type=detection --encode_type=jpg --encoded=true  --shuffle=true \
        --label_map_file=../../data/VOC0712/labelmap_emotor.prototxt \
        ../../data/VOC0712/VOCdevkit/ \
        ../../data/VOC0712/trainval.txt \
        ../../data/emotor_lmdb

轉換得到的 lmdb 文件放在 caffe-ssd/data/emotor_lmdb 下。

至此數據准備完畢。

准備訓練

我們用來進行遷移學習的基礎模型是 ssd_mobilenet_V1，其他預訓練模型可以去 caffe model zoo 找找。

ssd_mobilenet_V1 這個 github 倉庫里的腳本gen.py可用來創建訓練用的train.prototxt也可以使用templete下的模板來自行修改。

usage: gen.py [-h] [-s STAGE] [-d LMDB] [-l LABEL_MAP] [--classifier]
              [--size SIZE] -c CLASS_NUM

我的labelmap_emotor.prototxt里是兩個類，一個背景，一個是電動車，所以需要將輸出層的 num_class 設置為 2。

python gen.py -s=train -d=~/caffe-ssd/data/emotor_lmdb/ \
                -l=~/caffe-ssd/data/VOC0712/labelmap_emotor.prototxt \
                -c=2

產生的train.prototxt中些層的名字需要修改，那些需要修改可以看我的train.prototxt中的內容，所有后綴為_my的層都是被修改的。亦可以直接執行訓練，會有提示哪些層名字需要修改。

trainsform_param中的:

scale:0.007843
mean_value:127.5

這幾個參數在將模型用於部署的時候很關鍵。

name: "MobileNet-SSD"
layer {
  name: "data"
  type: "AnnotatedData"
  top: "data"
  top: "label"
  include {
    phase: TRAIN
  }
  transform_param {
    scale: 0.007843
    mirror: true
    mean_value: 127.5
    mean_value: 127.5
    mean_value: 127.5
    crop_size: 300
    resize_param {
      prob: 1.0
      resize_mode: WARP
      height: 300
      width: 300
      interp_mode: LINEAR
      interp_mode: AREA
      interp_mode: NEAREST
      interp_mode: CUBIC
      interp_mode: LANCZOS4
    }
    emit_constraint {
      emit_type: CENTER
    }
    distort_param {
      brightness_prob: 0.5
      brightness_delta: 32.0
      contrast_prob: 0.5
      contrast_lower: 0.5
      contrast_upper: 1.5
      hue_prob: 0.5
      hue_delta: 18.0
      saturation_prob: 0.5
      saturation_lower: 0.5
      saturation_upper: 1.5
      random_order_prob: 0.0
    }
    expand_param {
      prob: 0.5
      max_expand_ratio: 4.0
    }
  }
  data_param {
    source: "/root/caffe-ssd/data/emotor_lmdb"
    batch_size: 32
    backend: LMDB
  }
  annotated_data_param {
    batch_sampler {
      max_sample: 1
      max_trials: 1
    }
    batch_sampler {
      sampler {
        min_scale: 0.3
        max_scale: 1.0
        min_aspect_ratio: 0.5
        max_aspect_ratio: 2.0
      }
      sample_constraint {
        min_jaccard_overlap: 0.1
      }
      max_sample: 1
      max_trials: 50
    }
    batch_sampler {
      sampler {
        min_scale: 0.3
        max_scale: 1.0
        min_aspect_ratio: 0.5
        max_aspect_ratio: 2.0
      }
      sample_constraint {
        min_jaccard_overlap: 0.3
      }
      max_sample: 1
      max_trials: 50
    }
    batch_sampler {
      sampler {
        min_scale: 0.3
        max_scale: 1.0
        min_aspect_ratio: 0.5
        max_aspect_ratio: 2.0
      }
      sample_constraint {
        min_jaccard_overlap: 0.5
      }
      max_sample: 1
      max_trials: 50
    }
    batch_sampler {
      sampler {
        min_scale: 0.3
        max_scale: 1.0
        min_aspect_ratio: 0.5
        max_aspect_ratio: 2.0
      }
      sample_constraint {
        min_jaccard_overlap: 0.7
      }
      max_sample: 1
      max_trials: 50
    }
    batch_sampler {
      sampler {
        min_scale: 0.3
        max_scale: 1.0
        min_aspect_ratio: 0.5
        max_aspect_ratio: 2.0
      }
      sample_constraint {
        min_jaccard_overlap: 0.9
      }
      max_sample: 1
      max_trials: 50
    }
    batch_sampler {
      sampler {
        min_scale: 0.3
        max_scale: 1.0
        min_aspect_ratio: 0.5
        max_aspect_ratio: 2.0
      }
      sample_constraint {
        max_jaccard_overlap: 1.0
      }
      max_sample: 1
      max_trials: 50
    }
    label_map_file: "~/caffe-ssd/data/VOC0712/labelmap_emotor.prototxt"
  }
}
layer {
  name: "conv0"
  type: "Convolution"
  bottom: "data"
  top: "conv0"
  param {
    lr_mult: 0.1
    decay_mult: 0.1
  }
  convolution_param {
    num_output: 32
    bias_term: false
    pad: 1
    kernel_size: 3
    stride: 2
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv0/bn"
  type: "BatchNorm"
  bottom: "conv0"
  top: "conv0"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
}
layer {
  name: "conv0/scale"
  type: "Scale"
  bottom: "conv0"
  top: "conv0"
  param {
    lr_mult: 0.1
    decay_mult: 0.0
  }
  param {
    lr_mult: 0.2
    decay_mult: 0.0
  }
  scale_param {
    filler {
      value: 1
    }
    bias_term: true
    bias_filler {
      value: 0
    }
  }
}
layer {
  name: "conv0/relu"
  type: "ReLU"
  bottom: "conv0"
  top: "conv0"
}
layer {
  name: "conv1/dw"
  type: "Convolution"
  bottom: "conv0"
  top: "conv1/dw"
  param {
    lr_mult: 0.1
    decay_mult: 0.1
  }
  convolution_param {
    num_output: 32
    bias_term: false
    pad: 1
    kernel_size: 3
    group: 32
    engine: CAFFE
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv1/dw/bn"
  type: "BatchNorm"
  bottom: "conv1/dw"
  top: "conv1/dw"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
}
layer {
  name: "conv1/dw/scale"
  type: "Scale"
  bottom: "conv1/dw"
  top: "conv1/dw"
  param {
    lr_mult: 0.1
    decay_mult: 0.0
  }
  param {
    lr_mult: 0.2
    decay_mult: 0.0
  }
  scale_param {
    filler {
      value: 1
    }
    bias_term: true
    bias_filler {
      value: 0
    }
  }
}
layer {
  name: "conv1/dw/relu"
  type: "ReLU"
  bottom: "conv1/dw"
  top: "conv1/dw"
}
layer {
  name: "conv1"
  type: "Convolution"
  bottom: "conv1/dw"
  top: "conv1"
  param {
    lr_mult: 0.1
    decay_mult: 0.1
  }
  convolution_param {
    num_output: 64
    bias_term: false
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv1/bn"
  type: "BatchNorm"
  bottom: "conv1"
  top: "conv1"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
}
layer {
  name: "conv1/scale"
  type: "Scale"
  bottom: "conv1"
  top: "conv1"
  param {
    lr_mult: 0.1
    decay_mult: 0.0
  }
  param {
    lr_mult: 0.2
    decay_mult: 0.0
  }
  scale_param {
    filler {
      value: 1
    }
    bias_term: true
    bias_filler {
      value: 0
    }
  }
}
layer {
  name: "conv1/relu"
  type: "ReLU"
  bottom: "conv1"
  top: "conv1"
}
layer {
  name: "conv2/dw"
  type: "Convolution"
  bottom: "conv1"
  top: "conv2/dw"
  param {
    lr_mult: 0.1
    decay_mult: 0.1
  }
  convolution_param {
    num_output: 64
    bias_term: false
    pad: 1
    kernel_size: 3
    stride: 2
    group: 64
    engine: CAFFE
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv2/dw/bn"
  type: "BatchNorm"
  bottom: "conv2/dw"
  top: "conv2/dw"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
}
layer {
  name: "conv2/dw/scale"
  type: "Scale"
  bottom: "conv2/dw"
  top: "conv2/dw"
  param {
    lr_mult: 0.1
    decay_mult: 0.0
  }
  param {
    lr_mult: 0.2
    decay_mult: 0.0
  }
  scale_param {
    filler {
      value: 1
    }
    bias_term: true
    bias_filler {
      value: 0
    }
  }
}
layer {
  name: "conv2/dw/relu"
  type: "ReLU"
  bottom: "conv2/dw"
  top: "conv2/dw"
}
layer {
  name: "conv2"
  type: "Convolution"
  bottom: "conv2/dw"
  top: "conv2"
  param {
    lr_mult: 0.1
    decay_mult: 0.1
  }
  convolution_param {
    num_output: 128
    bias_term: false
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv2/bn"
  type: "BatchNorm"
  bottom: "conv2"
  top: "conv2"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
}
layer {
  name: "conv2/scale"
  type: "Scale"
  bottom: "conv2"
  top: "conv2"
  param {
    lr_mult: 0.1
    decay_mult: 0.0
  }
  param {
    lr_mult: 0.2
    decay_mult: 0.0
  }
  scale_param {
    filler {
      value: 1
    }
    bias_term: true
    bias_filler {
      value: 0
    }
  }
}
layer {
  name: "conv2/relu"
  type: "ReLU"
  bottom: "conv2"
  top: "conv2"
}
layer {
  name: "conv3/dw"
  type: "Convolution"
  bottom: "conv2"
  top: "conv3/dw"
  param {
    lr_mult: 0.1
    decay_mult: 0.1
  }
  convolution_param {
    num_output: 128
    bias_term: false
    pad: 1
    kernel_size: 3
    group: 128
    engine: CAFFE
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv3/dw/bn"
  type: "BatchNorm"
  bottom: "conv3/dw"
  top: "conv3/dw"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
}
layer {
  name: "conv3/dw/scale"
  type: "Scale"
  bottom: "conv3/dw"
  top: "conv3/dw"
  param {
    lr_mult: 0.1
    decay_mult: 0.0
  }
  param {
    lr_mult: 0.2
    decay_mult: 0.0
  }
  scale_param {
    filler {
      value: 1
    }
    bias_term: true
    bias_filler {
      value: 0
    }
  }
}
layer {
  name: "conv3/dw/relu"
  type: "ReLU"
  bottom: "conv3/dw"
  top: "conv3/dw"
}
layer {
  name: "conv3"
  type: "Convolution"
  bottom: "conv3/dw"
  top: "conv3"
  param {
    lr_mult: 0.1
    decay_mult: 0.1
  }
  convolution_param {
    num_output: 128
    bias_term: false
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv3/bn"
  type: "BatchNorm"
  bottom: "conv3"
  top: "conv3"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
}
layer {
  name: "conv3/scale"
  type: "Scale"
  bottom: "conv3"
  top: "conv3"
  param {
    lr_mult: 0.1
    decay_mult: 0.0
  }
  param {
    lr_mult: 0.2
    decay_mult: 0.0
  }
  scale_param {
    filler {
      value: 1
    }
    bias_term: true
    bias_filler {
      value: 0
    }
  }
}
layer {
  name: "conv3/relu"
  type: "ReLU"
  bottom: "conv3"
  top: "conv3"
}
layer {
  name: "conv4/dw"
  type: "Convolution"
  bottom: "conv3"
  top: "conv4/dw"
  param {
    lr_mult: 0.1
    decay_mult: 0.1
  }
  convolution_param {
    num_output: 128
    bias_term: false
    pad: 1
    kernel_size: 3
    stride: 2
    group: 128
    engine: CAFFE
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv4/dw/bn"
  type: "BatchNorm"
  bottom: "conv4/dw"
  top: "conv4/dw"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
}
layer {
  name: "conv4/dw/scale"
  type: "Scale"
  bottom: "conv4/dw"
  top: "conv4/dw"
  param {
    lr_mult: 0.1
    decay_mult: 0.0
  }
  param {
    lr_mult: 0.2
    decay_mult: 0.0
  }
  scale_param {
    filler {
      value: 1
    }
    bias_term: true
    bias_filler {
      value: 0
    }
  }
}
layer {
  name: "conv4/dw/relu"
  type: "ReLU"
  bottom: "conv4/dw"
  top: "conv4/dw"
}
layer {
  name: "conv4"
  type: "Convolution"
  bottom: "conv4/dw"
  top: "conv4"
  param {
    lr_mult: 0.1
    decay_mult: 0.1
  }
  convolution_param {
    num_output: 256
    bias_term: false
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv4/bn"
  type: "BatchNorm"
  bottom: "conv4"
  top: "conv4"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
}
layer {
  name: "conv4/scale"
  type: "Scale"
  bottom: "conv4"
  top: "conv4"
  param {
    lr_mult: 0.1
    decay_mult: 0.0
  }
  param {
    lr_mult: 0.2
    decay_mult: 0.0
  }
  scale_param {
    filler {
      value: 1
    }
    bias_term: true
    bias_filler {
      value: 0
    }
  }
}
layer {
  name: "conv4/relu"
  type: "ReLU"
  bottom: "conv4"
  top: "conv4"
}
layer {
  name: "conv5/dw"
  type: "Convolution"
  bottom: "conv4"
  top: "conv5/dw"
  param {
    lr_mult: 0.1
    decay_mult: 0.1
  }
  convolution_param {
    num_output: 256
    bias_term: false
    pad: 1
    kernel_size: 3
    group: 256
    engine: CAFFE
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv5/dw/bn"
  type: "BatchNorm"
  bottom: "conv5/dw"
  top: "conv5/dw"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
}
layer {
  name: "conv5/dw/scale"
  type: "Scale"
  bottom: "conv5/dw"
  top: "conv5/dw"
  param {
    lr_mult: 0.1
    decay_mult: 0.0
  }
  param {
    lr_mult: 0.2
    decay_mult: 0.0
  }
  scale_param {
    filler {
      value: 1
    }
    bias_term: true
    bias_filler {
      value: 0
    }
  }
}
layer {
  name: "conv5/dw/relu"
  type: "ReLU"
  bottom: "conv5/dw"
  top: "conv5/dw"
}
layer {
  name: "conv5"
  type: "Convolution"
  bottom: "conv5/dw"
  top: "conv5"
  param {
    lr_mult: 0.1
    decay_mult: 0.1
  }
  convolution_param {
    num_output: 256
    bias_term: false
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv5/bn"
  type: "BatchNorm"
  bottom: "conv5"
  top: "conv5"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
}
layer {
  name: "conv5/scale"
  type: "Scale"
  bottom: "conv5"
  top: "conv5"
  param {
    lr_mult: 0.1
    decay_mult: 0.0
  }
  param {
    lr_mult: 0.2
    decay_mult: 0.0
  }
  scale_param {
    filler {
      value: 1
    }
    bias_term: true
    bias_filler {
      value: 0
    }
  }
}
layer {
  name: "conv5/relu"
  type: "ReLU"
  bottom: "conv5"
  top: "conv5"
}
layer {
  name: "conv6/dw"
  type: "Convolution"
  bottom: "conv5"
  top: "conv6/dw"
  param {
    lr_mult: 0.1
    decay_mult: 0.1
  }
  convolution_param {
    num_output: 256
    bias_term: false
    pad: 1
    kernel_size: 3
    stride: 2
    group: 256
    engine: CAFFE
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv6/dw/bn"
  type: "BatchNorm"
  bottom: "conv6/dw"
  top: "conv6/dw"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
}
layer {
  name: "conv6/dw/scale"
  type: "Scale"
  bottom: "conv6/dw"
  top: "conv6/dw"
  param {
    lr_mult: 0.1
    decay_mult: 0.0
  }
  param {
    lr_mult: 0.2
    decay_mult: 0.0
  }
  scale_param {
    filler {
      value: 1
    }
    bias_term: true
    bias_filler {
      value: 0
    }
  }
}
layer {
  name: "conv6/dw/relu"
  type: "ReLU"
  bottom: "conv6/dw"
  top: "conv6/dw"
}
layer {
  name: "conv6"
  type: "Convolution"
  bottom: "conv6/dw"
  top: "conv6"
  param {
    lr_mult: 0.1
    decay_mult: 0.1
  }
  convolution_param {
    num_output: 512
    bias_term: false
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv6/bn"
  type: "BatchNorm"
  bottom: "conv6"
  top: "conv6"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
}
layer {
  name: "conv6/scale"
  type: "Scale"
  bottom: "conv6"
  top: "conv6"
  param {
    lr_mult: 0.1
    decay_mult: 0.0
  }
  param {
    lr_mult: 0.2
    decay_mult: 0.0
  }
  scale_param {
    filler {
      value: 1
    }
    bias_term: true
    bias_filler {
      value: 0
    }
  }
}
layer {
  name: "conv6/relu"
  type: "ReLU"
  bottom: "conv6"
  top: "conv6"
}
layer {
  name: "conv7/dw"
  type: "Convolution"
  bottom: "conv6"
  top: "conv7/dw"
  param {
    lr_mult: 0.1
    decay_mult: 0.1
  }
  convolution_param {
    num_output: 512
    bias_term: false
    pad: 1
    kernel_size: 3
    group: 512
    engine: CAFFE
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv7/dw/bn"
  type: "BatchNorm"
  bottom: "conv7/dw"
  top: "conv7/dw"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
}
layer {
  name: "conv7/dw/scale"
  type: "Scale"
  bottom: "conv7/dw"
  top: "conv7/dw"
  param {
    lr_mult: 0.1
    decay_mult: 0.0
  }
  param {
    lr_mult: 0.2
    decay_mult: 0.0
  }
  scale_param {
    filler {
      value: 1
    }
    bias_term: true
    bias_filler {
      value: 0
    }
  }
}
layer {
  name: "conv7/dw/relu"
  type: "ReLU"
  bottom: "conv7/dw"
  top: "conv7/dw"
}
layer {
  name: "conv7"
  type: "Convolution"
  bottom: "conv7/dw"
  top: "conv7"
  param {
    lr_mult: 0.1
    decay_mult: 0.1
  }
  convolution_param {
    num_output: 512
    bias_term: false
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv7/bn"
  type: "BatchNorm"
  bottom: "conv7"
  top: "conv7"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
}
layer {
  name: "conv7/scale"
  type: "Scale"
  bottom: "conv7"
  top: "conv7"
  param {
    lr_mult: 0.1
    decay_mult: 0.0
  }
  param {
    lr_mult: 0.2
    decay_mult: 0.0
  }
  scale_param {
    filler {
      value: 1
    }
    bias_term: true
    bias_filler {
      value: 0
    }
  }
}
layer {
  name: "conv7/relu"
  type: "ReLU"
  bottom: "conv7"
  top: "conv7"
}
layer {
  name: "conv8/dw"
  type: "Convolution"
  bottom: "conv7"
  top: "conv8/dw"
  param {
    lr_mult: 0.1
    decay_mult: 0.1
  }
  convolution_param {
    num_output: 512
    bias_term: false
    pad: 1
    kernel_size: 3
    group: 512
    engine: CAFFE
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv8/dw/bn"
  type: "BatchNorm"
  bottom: "conv8/dw"
  top: "conv8/dw"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
}
layer {
  name: "conv8/dw/scale"
  type: "Scale"
  bottom: "conv8/dw"
  top: "conv8/dw"
  param {
    lr_mult: 0.1
    decay_mult: 0.0
  }
  param {
    lr_mult: 0.2
    decay_mult: 0.0
  }
  scale_param {
    filler {
      value: 1
    }
    bias_term: true
    bias_filler {
      value: 0
    }
  }
}
layer {
  name: "conv8/dw/relu"
  type: "ReLU"
  bottom: "conv8/dw"
  top: "conv8/dw"
}
layer {
  name: "conv8"
  type: "Convolution"
  bottom: "conv8/dw"
  top: "conv8"
  param {
    lr_mult: 0.1
    decay_mult: 0.1
  }
  convolution_param {
    num_output: 512
    bias_term: false
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv8/bn"
  type: "BatchNorm"
  bottom: "conv8"
  top: "conv8"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
}
layer {
  name: "conv8/scale"
  type: "Scale"
  bottom: "conv8"
  top: "conv8"
  param {
    lr_mult: 0.1
    decay_mult: 0.0
  }
  param {
    lr_mult: 0.2
    decay_mult: 0.0
  }
  scale_param {
    filler {
      value: 1
    }
    bias_term: true
    bias_filler {
      value: 0
    }
  }
}
layer {
  name: "conv8/relu"
  type: "ReLU"
  bottom: "conv8"
  top: "conv8"
}
layer {
  name: "conv9/dw"
  type: "Convolution"
  bottom: "conv8"
  top: "conv9/dw"
  param {
    lr_mult: 0.1
    decay_mult: 0.1
  }
  convolution_param {
    num_output: 512
    bias_term: false
    pad: 1
    kernel_size: 3
    group: 512
    engine: CAFFE
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv9/dw/bn"
  type: "BatchNorm"
  bottom: "conv9/dw"
  top: "conv9/dw"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
}
layer {
  name: "conv9/dw/scale"
  type: "Scale"
  bottom: "conv9/dw"
  top: "conv9/dw"
  param {
    lr_mult: 0.1
    decay_mult: 0.0
  }
  param {
    lr_mult: 0.2
    decay_mult: 0.0
  }
  scale_param {
    filler {
      value: 1
    }
    bias_term: true
    bias_filler {
      value: 0
    }
  }
}
layer {
  name: "conv9/dw/relu"
  type: "ReLU"
  bottom: "conv9/dw"
  top: "conv9/dw"
}
layer {
  name: "conv9"
  type: "Convolution"
  bottom: "conv9/dw"
  top: "conv9"
  param {
    lr_mult: 0.1
    decay_mult: 0.1
  }
  convolution_param {
    num_output: 512
    bias_term: false
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv9/bn"
  type: "BatchNorm"
  bottom: "conv9"
  top: "conv9"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
}
layer {
  name: "conv9/scale"
  type: "Scale"
  bottom: "conv9"
  top: "conv9"
  param {
    lr_mult: 0.1
    decay_mult: 0.0
  }
  param {
    lr_mult: 0.2
    decay_mult: 0.0
  }
  scale_param {
    filler {
      value: 1
    }
    bias_term: true
    bias_filler {
      value: 0
    }
  }
}
layer {
  name: "conv9/relu"
  type: "ReLU"
  bottom: "conv9"
  top: "conv9"
}
layer {
  name: "conv10/dw"
  type: "Convolution"
  bottom: "conv9"
  top: "conv10/dw"
  param {
    lr_mult: 0.1
    decay_mult: 0.1
  }
  convolution_param {
    num_output: 512
    bias_term: false
    pad: 1
    kernel_size: 3
    group: 512
    engine: CAFFE
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv10/dw/bn"
  type: "BatchNorm"
  bottom: "conv10/dw"
  top: "conv10/dw"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
}
layer {
  name: "conv10/dw/scale"
  type: "Scale"
  bottom: "conv10/dw"
  top: "conv10/dw"
  param {
    lr_mult: 0.1
    decay_mult: 0.0
  }
  param {
    lr_mult: 0.2
    decay_mult: 0.0
  }
  scale_param {
    filler {
      value: 1
    }
    bias_term: true
    bias_filler {
      value: 0
    }
  }
}
layer {
  name: "conv10/dw/relu"
  type: "ReLU"
  bottom: "conv10/dw"
  top: "conv10/dw"
}
layer {
  name: "conv10"
  type: "Convolution"
  bottom: "conv10/dw"
  top: "conv10"
  param {
    lr_mult: 0.1
    decay_mult: 0.1
  }
  convolution_param {
    num_output: 512
    bias_term: false
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv10/bn"
  type: "BatchNorm"
  bottom: "conv10"
  top: "conv10"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
}
layer {
  name: "conv10/scale"
  type: "Scale"
  bottom: "conv10"
  top: "conv10"
  param {
    lr_mult: 0.1
    decay_mult: 0.0
  }
  param {
    lr_mult: 0.2
    decay_mult: 0.0
  }
  scale_param {
    filler {
      value: 1
    }
    bias_term: true
    bias_filler {
      value: 0
    }
  }
}
layer {
  name: "conv10/relu"
  type: "ReLU"
  bottom: "conv10"
  top: "conv10"
}
layer {
  name: "conv11/dw"
  type: "Convolution"
  bottom: "conv10"
  top: "conv11/dw"
  param {
    lr_mult: 0.1
    decay_mult: 0.1
  }
  convolution_param {
    num_output: 512
    bias_term: false
    pad: 1
    kernel_size: 3
    group: 512
    engine: CAFFE
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv11/dw/bn"
  type: "BatchNorm"
  bottom: "conv11/dw"
  top: "conv11/dw"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
}
layer {
  name: "conv11/dw/scale"
  type: "Scale"
  bottom: "conv11/dw"
  top: "conv11/dw"
  param {
    lr_mult: 0.1
    decay_mult: 0.0
  }
  param {
    lr_mult: 0.2
    decay_mult: 0.0
  }
  scale_param {
    filler {
      value: 1
    }
    bias_term: true
    bias_filler {
      value: 0
    }
  }
}
layer {
  name: "conv11/dw/relu"
  type: "ReLU"
  bottom: "conv11/dw"
  top: "conv11/dw"
}
layer {
  name: "conv11"
  type: "Convolution"
  bottom: "conv11/dw"
  top: "conv11"
  param {
    lr_mult: 0.1
    decay_mult: 0.1
  }
  convolution_param {
    num_output: 512
    bias_term: false
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv11/bn"
  type: "BatchNorm"
  bottom: "conv11"
  top: "conv11"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
}
layer {
  name: "conv11/scale"
  type: "Scale"
  bottom: "conv11"
  top: "conv11"
  param {
    lr_mult: 0.1
    decay_mult: 0.0
  }
  param {
    lr_mult: 0.2
    decay_mult: 0.0
  }
  scale_param {
    filler {
      value: 1
    }
    bias_term: true
    bias_filler {
      value: 0
    }
  }
}
layer {
  name: "conv11/relu"
  type: "ReLU"
  bottom: "conv11"
  top: "conv11"
}
layer {
  name: "conv12/dw"
  type: "Convolution"
  bottom: "conv11"
  top: "conv12/dw"
  param {
    lr_mult: 0.1
    decay_mult: 0.1
  }
  convolution_param {
    num_output: 512
    bias_term: false
    pad: 1
    kernel_size: 3
    stride: 2
    group: 512
    engine: CAFFE
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv12/dw/bn"
  type: "BatchNorm"
  bottom: "conv12/dw"
  top: "conv12/dw"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
}
layer {
  name: "conv12/dw/scale"
  type: "Scale"
  bottom: "conv12/dw"
  top: "conv12/dw"
  param {
    lr_mult: 0.1
    decay_mult: 0.0
  }
  param {
    lr_mult: 0.2
    decay_mult: 0.0
  }
  scale_param {
    filler {
      value: 1
    }
    bias_term: true
    bias_filler {
      value: 0
    }
  }
}
layer {
  name: "conv12/dw/relu"
  type: "ReLU"
  bottom: "conv12/dw"
  top: "conv12/dw"
}
layer {
  name: "conv12"
  type: "Convolution"
  bottom: "conv12/dw"
  top: "conv12"
  param {
    lr_mult: 0.1
    decay_mult: 0.1
  }
  convolution_param {
    num_output: 1024
    bias_term: false
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv12/bn"
  type: "BatchNorm"
  bottom: "conv12"
  top: "conv12"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
}
layer {
  name: "conv12/scale"
  type: "Scale"
  bottom: "conv12"
  top: "conv12"
  param {
    lr_mult: 0.1
    decay_mult: 0.0
  }
  param {
    lr_mult: 0.2
    decay_mult: 0.0
  }
  scale_param {
    filler {
      value: 1
    }
    bias_term: true
    bias_filler {
      value: 0
    }
  }
}
layer {
  name: "conv12/relu"
  type: "ReLU"
  bottom: "conv12"
  top: "conv12"
}
layer {
  name: "conv13/dw"
  type: "Convolution"
  bottom: "conv12"
  top: "conv13/dw"
  param {
    lr_mult: 0.1
    decay_mult: 0.1
  }
  convolution_param {
    num_output: 1024
    bias_term: false
    pad: 1
    kernel_size: 3
    group: 1024
    engine: CAFFE
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv13/dw/bn"
  type: "BatchNorm"
  bottom: "conv13/dw"
  top: "conv13/dw"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
}
layer {
  name: "conv13/dw/scale"
  type: "Scale"
  bottom: "conv13/dw"
  top: "conv13/dw"
  param {
    lr_mult: 0.1
    decay_mult: 0.0
  }
  param {
    lr_mult: 0.2
    decay_mult: 0.0
  }
  scale_param {
    filler {
      value: 1
    }
    bias_term: true
    bias_filler {
      value: 0
    }
  }
}
layer {
  name: "conv13/dw/relu"
  type: "ReLU"
  bottom: "conv13/dw"
  top: "conv13/dw"
}
layer {
  name: "conv13"
  type: "Convolution"
  bottom: "conv13/dw"
  top: "conv13"
  param {
    lr_mult: 0.1
    decay_mult: 0.1
  }
  convolution_param {
    num_output: 1024
    bias_term: false
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv13/bn"
  type: "BatchNorm"
  bottom: "conv13"
  top: "conv13"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
}
layer {
  name: "conv13/scale"
  type: "Scale"
  bottom: "conv13"
  top: "conv13"
  param {
    lr_mult: 0.1
    decay_mult: 0.0
  }
  param {
    lr_mult: 0.2
    decay_mult: 0.0
  }
  scale_param {
    filler {
      value: 1
    }
    bias_term: true
    bias_filler {
      value: 0
    }
  }
}
layer {
  name: "conv13/relu"
  type: "ReLU"
  bottom: "conv13"
  top: "conv13"
}
layer {
  name: "conv14_1"
  type: "Convolution"
  bottom: "conv13"
  top: "conv14_1"
  param {
    lr_mult: 0.1
    decay_mult: 0.1
  }
  convolution_param {
    num_output: 256
    bias_term: false
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv14_1/bn"
  type: "BatchNorm"
  bottom: "conv14_1"
  top: "conv14_1"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
}
layer {
  name: "conv14_1/scale"
  type: "Scale"
  bottom: "conv14_1"
  top: "conv14_1"
  param {
    lr_mult: 0.1
    decay_mult: 0.0
  }
  param {
    lr_mult: 0.2
    decay_mult: 0.0
  }
  scale_param {
    filler {
      value: 1
    }
    bias_term: true
    bias_filler {
      value: 0
    }
  }
}
layer {
  name: "conv14_1/relu"
  type: "ReLU"
  bottom: "conv14_1"
  top: "conv14_1"
}
layer {
  name: "conv14_2"
  type: "Convolution"
  bottom: "conv14_1"
  top: "conv14_2"
  param {
    lr_mult: 0.1
    decay_mult: 0.1
  }
  convolution_param {
    num_output: 512
    bias_term: false
    pad: 1
    kernel_size: 3
    stride: 2
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv14_2/bn"
  type: "BatchNorm"
  bottom: "conv14_2"
  top: "conv14_2"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
}
layer {
  name: "conv14_2/scale"
  type: "Scale"
  bottom: "conv14_2"
  top: "conv14_2"
  param {
    lr_mult: 0.1
    decay_mult: 0.0
  }
  param {
    lr_mult: 0.2
    decay_mult: 0.0
  }
  scale_param {
    filler {
      value: 1
    }
    bias_term: true
    bias_filler {
      value: 0
    }
  }
}
layer {
  name: "conv14_2/relu"
  type: "ReLU"
  bottom: "conv14_2"
  top: "conv14_2"
}
layer {
  name: "conv15_1"
  type: "Convolution"
  bottom: "conv14_2"
  top: "conv15_1"
  param {
    lr_mult: 0.1
    decay_mult: 0.1
  }
  convolution_param {
    num_output: 128
    bias_term: false
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv15_1/bn"
  type: "BatchNorm"
  bottom: "conv15_1"
  top: "conv15_1"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
}
layer {
  name: "conv15_1/scale"
  type: "Scale"
  bottom: "conv15_1"
  top: "conv15_1"
  param {
    lr_mult: 0.1
    decay_mult: 0.0
  }
  param {
    lr_mult: 0.2
    decay_mult: 0.0
  }
  scale_param {
    filler {
      value: 1
    }
    bias_term: true
    bias_filler {
      value: 0
    }
  }
}
layer {
  name: "conv15_1/relu"
  type: "ReLU"
  bottom: "conv15_1"
  top: "conv15_1"
}
layer {
  name: "conv15_2"
  type: "Convolution"
  bottom: "conv15_1"
  top: "conv15_2"
  param {
    lr_mult: 0.1
    decay_mult: 0.1
  }
  convolution_param {
    num_output: 256
    bias_term: false
    pad: 1
    kernel_size: 3
    stride: 2
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv15_2/bn"
  type: "BatchNorm"
  bottom: "conv15_2"
  top: "conv15_2"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
}
layer {
  name: "conv15_2/scale"
  type: "Scale"
  bottom: "conv15_2"
  top: "conv15_2"
  param {
    lr_mult: 0.1
    decay_mult: 0.0
  }
  param {
    lr_mult: 0.2
    decay_mult: 0.0
  }
  scale_param {
    filler {
      value: 1
    }
    bias_term: true
    bias_filler {
      value: 0
    }
  }
}
layer {
  name: "conv15_2/relu"
  type: "ReLU"
  bottom: "conv15_2"
  top: "conv15_2"
}
layer {
  name: "conv16_1"
  type: "Convolution"
  bottom: "conv15_2"
  top: "conv16_1"
  param {
    lr_mult: 0.1
    decay_mult: 0.1
  }
  convolution_param {
    num_output: 128
    bias_term: false
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv16_1/bn"
  type: "BatchNorm"
  bottom: "conv16_1"
  top: "conv16_1"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
}
layer {
  name: "conv16_1/scale"
  type: "Scale"
  bottom: "conv16_1"
  top: "conv16_1"
  param {
    lr_mult: 0.1
    decay_mult: 0.0
  }
  param {
    lr_mult: 0.2
    decay_mult: 0.0
  }
  scale_param {
    filler {
      value: 1
    }
    bias_term: true
    bias_filler {
      value: 0
    }
  }
}
layer {
  name: "conv16_1/relu"
  type: "ReLU"
  bottom: "conv16_1"
  top: "conv16_1"
}
layer {
  name: "conv16_2"
  type: "Convolution"
  bottom: "conv16_1"
  top: "conv16_2"
  param {
    lr_mult: 0.1
    decay_mult: 0.1
  }
  convolution_param {
    num_output: 256
    bias_term: false
    pad: 1
    kernel_size: 3
    stride: 2
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv16_2/bn"
  type: "BatchNorm"
  bottom: "conv16_2"
  top: "conv16_2"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
}
layer {
  name: "conv16_2/scale"
  type: "Scale"
  bottom: "conv16_2"
  top: "conv16_2"
  param {
    lr_mult: 0.1
    decay_mult: 0.0
  }
  param {
    lr_mult: 0.2
    decay_mult: 0.0
  }
  scale_param {
    filler {
      value: 1
    }
    bias_term: true
    bias_filler {
      value: 0
    }
  }
}
layer {
  name: "conv16_2/relu"
  type: "ReLU"
  bottom: "conv16_2"
  top: "conv16_2"
}
layer {
  name: "conv17_1"
  type: "Convolution"
  bottom: "conv16_2"
  top: "conv17_1"
  param {
    lr_mult: 0.1
    decay_mult: 0.1
  }
  convolution_param {
    num_output: 64
    bias_term: false
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv17_1/bn"
  type: "BatchNorm"
  bottom: "conv17_1"
  top: "conv17_1"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
}
layer {
  name: "conv17_1/scale"
  type: "Scale"
  bottom: "conv17_1"
  top: "conv17_1"
  param {
    lr_mult: 0.1
    decay_mult: 0.0
  }
  param {
    lr_mult: 0.2
    decay_mult: 0.0
  }
  scale_param {
    filler {
      value: 1
    }
    bias_term: true
    bias_filler {
      value: 0
    }
  }
}
layer {
  name: "conv17_1/relu"
  type: "ReLU"
  bottom: "conv17_1"
  top: "conv17_1"
}
layer {
  name: "conv17_2"
  type: "Convolution"
  bottom: "conv17_1"
  top: "conv17_2"
  param {
    lr_mult: 0.1
    decay_mult: 0.1
  }
  convolution_param {
    num_output: 128
    bias_term: false
    pad: 1
    kernel_size: 3
    stride: 2
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv17_2/bn"
  type: "BatchNorm"
  bottom: "conv17_2"
  top: "conv17_2"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
}
layer {
  name: "conv17_2/scale"
  type: "Scale"
  bottom: "conv17_2"
  top: "conv17_2"
  param {
    lr_mult: 0.1
    decay_mult: 0.0
  }
  param {
    lr_mult: 0.2
    decay_mult: 0.0
  }
  scale_param {
    filler {
      value: 1
    }
    bias_term: true
    bias_filler {
      value: 0
    }
  }
}
layer {
  name: "conv17_2/relu"
  type: "ReLU"
  bottom: "conv17_2"
  top: "conv17_2"
}
layer {
  name: "conv11_mbox_loc"
  type: "Convolution"
  bottom: "conv11"
  top: "conv11_mbox_loc"
  param {
    lr_mult: 0.1
    decay_mult: 0.1
  }
  param {
    lr_mult: 0.2
    decay_mult: 0.0
  }
  convolution_param {
    num_output: 12
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0.0
    }
  }
}
layer {
  name: "conv11_mbox_loc_perm"
  type: "Permute"
  bottom: "conv11_mbox_loc"
  top: "conv11_mbox_loc_perm"
  permute_param {
    order: 0
    order: 2
    order: 3
    order: 1
  }
}
layer {
  name: "conv11_mbox_loc_flat"
  type: "Flatten"
  bottom: "conv11_mbox_loc_perm"
  top: "conv11_mbox_loc_flat"
  flatten_param {
    axis: 1
  }
}
layer {
  name: "conv11_mbox_conf_my"
  type: "Convolution"
  bottom: "conv11"
  top: "conv11_mbox_conf"
  param {
    lr_mult: 1.0
    decay_mult: 1.0
  }
  param {
    lr_mult: 2.0
    decay_mult: 0.0
  }
  convolution_param {
    num_output: 6
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0.0
    }
  }
}
layer {
  name: "conv11_mbox_conf_perm"
  type: "Permute"
  bottom: "conv11_mbox_conf"
  top: "conv11_mbox_conf_perm"
  permute_param {
    order: 0
    order: 2
    order: 3
    order: 1
  }
}
layer {
  name: "conv11_mbox_conf_flat"
  type: "Flatten"
  bottom: "conv11_mbox_conf_perm"
  top: "conv11_mbox_conf_flat"
  flatten_param {
    axis: 1
  }
}
layer {
  name: "conv11_mbox_priorbox"
  type: "PriorBox"
  bottom: "conv11"
  bottom: "data"
  top: "conv11_mbox_priorbox"
  prior_box_param {
    min_size: 60.0
    aspect_ratio: 2.0
    flip: true
    clip: false
    variance: 0.1
    variance: 0.1
    variance: 0.2
    variance: 0.2
    offset: 0.5
  }
}
layer {
  name: "conv13_mbox_loc"
  type: "Convolution"
  bottom: "conv13"
  top: "conv13_mbox_loc"
  param {
    lr_mult: 0.1
    decay_mult: 0.1
  }
  param {
    lr_mult: 0.2
    decay_mult: 0.0
  }
  convolution_param {
    num_output: 24
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0.0
    }
  }
}
layer {
  name: "conv13_mbox_loc_perm"
  type: "Permute"
  bottom: "conv13_mbox_loc"
  top: "conv13_mbox_loc_perm"
  permute_param {
    order: 0
    order: 2
    order: 3
    order: 1
  }
}
layer {
  name: "conv13_mbox_loc_flat"
  type: "Flatten"
  bottom: "conv13_mbox_loc_perm"
  top: "conv13_mbox_loc_flat"
  flatten_param {
    axis: 1
  }
}
layer {
  name: "conv13_mbox_conf_my"
  type: "Convolution"
  bottom: "conv13"
  top: "conv13_mbox_conf"
  param {
    lr_mult: 1.0
    decay_mult: 1.0
  }
  param {
    lr_mult: 2.0
    decay_mult: 0.0
  }
  convolution_param {
    num_output: 12
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0.0
    }
  }
}
layer {
  name: "conv13_mbox_conf_perm"
  type: "Permute"
  bottom: "conv13_mbox_conf"
  top: "conv13_mbox_conf_perm"
  permute_param {
    order: 0
    order: 2
    order: 3
    order: 1
  }
}
layer {
  name: "conv13_mbox_conf_flat"
  type: "Flatten"
  bottom: "conv13_mbox_conf_perm"
  top: "conv13_mbox_conf_flat"
  flatten_param {
    axis: 1
  }
}
layer {
  name: "conv13_mbox_priorbox"
  type: "PriorBox"
  bottom: "conv13"
  bottom: "data"
  top: "conv13_mbox_priorbox"
  prior_box_param {
    min_size: 105.0
    max_size: 150.0
    aspect_ratio: 2.0
    aspect_ratio: 3.0
    flip: true
    clip: false
    variance: 0.1
    variance: 0.1
    variance: 0.2
    variance: 0.2
    offset: 0.5
  }
}
layer {
  name: "conv14_2_mbox_loc"
  type: "Convolution"
  bottom: "conv14_2"
  top: "conv14_2_mbox_loc"
  param {
    lr_mult: 0.1
    decay_mult: 0.1
  }
  param {
    lr_mult: 0.2
    decay_mult: 0.0
  }
  convolution_param {
    num_output: 24
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0.0
    }
  }
}
layer {
  name: "conv14_2_mbox_loc_perm"
  type: "Permute"
  bottom: "conv14_2_mbox_loc"
  top: "conv14_2_mbox_loc_perm"
  permute_param {
    order: 0
    order: 2
    order: 3
    order: 1
  }
}
layer {
  name: "conv14_2_mbox_loc_flat"
  type: "Flatten"
  bottom: "conv14_2_mbox_loc_perm"
  top: "conv14_2_mbox_loc_flat"
  flatten_param {
    axis: 1
  }
}
layer {
  name: "conv14_2_mbox_conf_my"
  type: "Convolution"
  bottom: "conv14_2"
  top: "conv14_2_mbox_conf"
  param {
    lr_mult: 1.0
    decay_mult: 1.0
  }
  param {
    lr_mult: 2.0
    decay_mult: 0.0
  }
  convolution_param {
    num_output: 12
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0.0
    }
  }
}
layer {
  name: "conv14_2_mbox_conf_perm"
  type: "Permute"
  bottom: "conv14_2_mbox_conf"
  top: "conv14_2_mbox_conf_perm"
  permute_param {
    order: 0
    order: 2
    order: 3
    order: 1
  }
}
layer {
  name: "conv14_2_mbox_conf_flat"
  type: "Flatten"
  bottom: "conv14_2_mbox_conf_perm"
  top: "conv14_2_mbox_conf_flat"
  flatten_param {
    axis: 1
  }
}
layer {
  name: "conv14_2_mbox_priorbox"
  type: "PriorBox"
  bottom: "conv14_2"
  bottom: "data"
  top: "conv14_2_mbox_priorbox"
  prior_box_param {
    min_size: 150.0
    max_size: 195.0
    aspect_ratio: 2.0
    aspect_ratio: 3.0
    flip: true
    clip: false
    variance: 0.1
    variance: 0.1
    variance: 0.2
    variance: 0.2
    offset: 0.5
  }
}
layer {
  name: "conv15_2_mbox_loc"
  type: "Convolution"
  bottom: "conv15_2"
  top: "conv15_2_mbox_loc"
  param {
    lr_mult: 0.1
    decay_mult: 0.1
  }
  param {
    lr_mult: 0.2
    decay_mult: 0.0
  }
  convolution_param {
    num_output: 24
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0.0
    }
  }
}
layer {
  name: "conv15_2_mbox_loc_perm"
  type: "Permute"
  bottom: "conv15_2_mbox_loc"
  top: "conv15_2_mbox_loc_perm"
  permute_param {
    order: 0
    order: 2
    order: 3
    order: 1
  }
}
layer {
  name: "conv15_2_mbox_loc_flat"
  type: "Flatten"
  bottom: "conv15_2_mbox_loc_perm"
  top: "conv15_2_mbox_loc_flat"
  flatten_param {
    axis: 1
  }
}
layer {
  name: "conv15_2_mbox_conf_my"
  type: "Convolution"
  bottom: "conv15_2"
  top: "conv15_2_mbox_conf"
  param {
    lr_mult: 1.0
    decay_mult: 1.0
  }
  param {
    lr_mult: 2.0
    decay_mult: 0.0
  }
  convolution_param {
    num_output: 12
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0.0
    }
  }
}
layer {
  name: "conv15_2_mbox_conf_perm"
  type: "Permute"
  bottom: "conv15_2_mbox_conf"
  top: "conv15_2_mbox_conf_perm"
  permute_param {
    order: 0
    order: 2
    order: 3
    order: 1
  }
}
layer {
  name: "conv15_2_mbox_conf_flat"
  type: "Flatten"
  bottom: "conv15_2_mbox_conf_perm"
  top: "conv15_2_mbox_conf_flat"
  flatten_param {
    axis: 1
  }
}
layer {
  name: "conv15_2_mbox_priorbox"
  type: "PriorBox"
  bottom: "conv15_2"
  bottom: "data"
  top: "conv15_2_mbox_priorbox"
  prior_box_param {
    min_size: 195.0
    max_size: 240.0
    aspect_ratio: 2.0
    aspect_ratio: 3.0
    flip: true
    clip: false
    variance: 0.1
    variance: 0.1
    variance: 0.2
    variance: 0.2
    offset: 0.5
  }
}
layer {
  name: "conv16_2_mbox_loc"
  type: "Convolution"
  bottom: "conv16_2"
  top: "conv16_2_mbox_loc"
  param {
    lr_mult: 0.1
    decay_mult: 0.1
  }
  param {
    lr_mult: 0.2
    decay_mult: 0.0
  }
  convolution_param {
    num_output: 24
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0.0
    }
  }
}
layer {
  name: "conv16_2_mbox_loc_perm"
  type: "Permute"
  bottom: "conv16_2_mbox_loc"
  top: "conv16_2_mbox_loc_perm"
  permute_param {
    order: 0
    order: 2
    order: 3
    order: 1
  }
}
layer {
  name: "conv16_2_mbox_loc_flat"
  type: "Flatten"
  bottom: "conv16_2_mbox_loc_perm"
  top: "conv16_2_mbox_loc_flat"
  flatten_param {
    axis: 1
  }
}
layer {
  name: "conv16_2_mbox_conf_my"
  type: "Convolution"
  bottom: "conv16_2"
  top: "conv16_2_mbox_conf"
  param {
    lr_mult: 1.0
    decay_mult: 1.0
  }
  param {
    lr_mult: 2.0
    decay_mult: 0.0
  }
  convolution_param {
    num_output: 12
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0.0
    }
  }
}
layer {
  name: "conv16_2_mbox_conf_perm"
  type: "Permute"
  bottom: "conv16_2_mbox_conf"
  top: "conv16_2_mbox_conf_perm"
  permute_param {
    order: 0
    order: 2
    order: 3
    order: 1
  }
}
layer {
  name: "conv16_2_mbox_conf_flat"
  type: "Flatten"
  bottom: "conv16_2_mbox_conf_perm"
  top: "conv16_2_mbox_conf_flat"
  flatten_param {
    axis: 1
  }
}
layer {
  name: "conv16_2_mbox_priorbox"
  type: "PriorBox"
  bottom: "conv16_2"
  bottom: "data"
  top: "conv16_2_mbox_priorbox"
  prior_box_param {
    min_size: 240.0
    max_size: 285.0
    aspect_ratio: 2.0
    aspect_ratio: 3.0
    flip: true
    clip: false
    variance: 0.1
    variance: 0.1
    variance: 0.2
    variance: 0.2
    offset: 0.5
  }
}
layer {
  name: "conv17_2_mbox_loc"
  type: "Convolution"
  bottom: "conv17_2"
  top: "conv17_2_mbox_loc"
  param {
    lr_mult: 0.1
    decay_mult: 0.1
  }
  param {
    lr_mult: 0.2
    decay_mult: 0.0
  }
  convolution_param {
    num_output: 24
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0.0
    }
  }
}
layer {
  name: "conv17_2_mbox_loc_perm"
  type: "Permute"
  bottom: "conv17_2_mbox_loc"
  top: "conv17_2_mbox_loc_perm"
  permute_param {
    order: 0
    order: 2
    order: 3
    order: 1
  }
}
layer {
  name: "conv17_2_mbox_loc_flat"
  type: "Flatten"
  bottom: "conv17_2_mbox_loc_perm"
  top: "conv17_2_mbox_loc_flat"
  flatten_param {
    axis: 1
  }
}
layer {
  name: "conv17_2_mbox_conf_my"
  type: "Convolution"
  bottom: "conv17_2"
  top: "conv17_2_mbox_conf"
  param {
    lr_mult: 1.0
    decay_mult: 1.0
  }
  param {
    lr_mult: 2.0
    decay_mult: 0.0
  }
  convolution_param {
    num_output: 12
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0.0
    }
  }
}
layer {
  name: "conv17_2_mbox_conf_perm"
  type: "Permute"
  bottom: "conv17_2_mbox_conf"
  top: "conv17_2_mbox_conf_perm"
  permute_param {
    order: 0
    order: 2
    order: 3
    order: 1
  }
}
layer {
  name: "conv17_2_mbox_conf_flat"
  type: "Flatten"
  bottom: "conv17_2_mbox_conf_perm"
  top: "conv17_2_mbox_conf_flat"
  flatten_param {
    axis: 1
  }
}
layer {
  name: "conv17_2_mbox_priorbox"
  type: "PriorBox"
  bottom: "conv17_2"
  bottom: "data"
  top: "conv17_2_mbox_priorbox"
  prior_box_param {
    min_size: 285.0
    max_size: 300.0
    aspect_ratio: 2.0
    aspect_ratio: 3.0
    flip: true
    clip: false
    variance: 0.1
    variance: 0.1
    variance: 0.2
    variance: 0.2
    offset: 0.5
  }
}
layer {
  name: "mbox_loc"
  type: "Concat"
  bottom: "conv11_mbox_loc_flat"
  bottom: "conv13_mbox_loc_flat"
  bottom: "conv14_2_mbox_loc_flat"
  bottom: "conv15_2_mbox_loc_flat"
  bottom: "conv16_2_mbox_loc_flat"
  bottom: "conv17_2_mbox_loc_flat"
  top: "mbox_loc"
  concat_param {
    axis: 1
  }
}
layer {
  name: "mbox_conf"
  type: "Concat"
  bottom: "conv11_mbox_conf_flat"
  bottom: "conv13_mbox_conf_flat"
  bottom: "conv14_2_mbox_conf_flat"
  bottom: "conv15_2_mbox_conf_flat"
  bottom: "conv16_2_mbox_conf_flat"
  bottom: "conv17_2_mbox_conf_flat"
  top: "mbox_conf"
  concat_param {
    axis: 1
  }
}
layer {
  name: "mbox_priorbox"
  type: "Concat"
  bottom: "conv11_mbox_priorbox"
  bottom: "conv13_mbox_priorbox"
  bottom: "conv14_2_mbox_priorbox"
  bottom: "conv15_2_mbox_priorbox"
  bottom: "conv16_2_mbox_priorbox"
  bottom: "conv17_2_mbox_priorbox"
  top: "mbox_priorbox"
  concat_param {
    axis: 2
  }
}
layer {
  name: "mbox_loss"
  type: "MultiBoxLoss"
  bottom: "mbox_loc"
  bottom: "mbox_conf"
  bottom: "mbox_priorbox"
  bottom: "label"
  top: "mbox_loss"
  include {
    phase: TRAIN
  }
  propagate_down: true
  propagate_down: true
  propagate_down: false
  propagate_down: false
  loss_param {
    normalization: VALID
  }
  multibox_loss_param {
    loc_loss_type: SMOOTH_L1
    conf_loss_type: SOFTMAX
    loc_weight: 1.0
    num_classes: 2
    share_location: true
    match_type: PER_PREDICTION
    overlap_threshold: 0.5
    use_prior_for_matching: true
    background_label_id: 0
    use_difficult_gt: true
    neg_pos_ratio: 3.0
    neg_overlap: 0.5
    code_type: CENTER_SIZE
    ignore_cross_boundary_bbox: false
    mining_type: MAX_NEGATIVE
  }
}

進行訓練

訓練時執行的命令：

./../../build/tools/caffe train --solver=./MobileNet-SSD/solver_train.prototxt \
                        --weights=./MobileNet-SSD/mobilenet_iter_73000.caffemodel

其中的solver_train.prototxt內容如下：

train_net: "/root/caffe-ssd/workplace/ssd_mobilenetv2/MobileNet-SSD/train.prototxt"
base_lr: 0.0005
display: 10
max_iter: 120000
lr_policy: "multistep"
gamma: 0.5
weight_decay: 0.00005
snapshot: 1000
snapshot_prefix: "/root/caffe-ssd/workplace/ssd_mobilenetv2/train"
solver_mode: GPU
debug_info: false
snapshot_after_train: true
test_initialization: false
average_loss: 10
stepvalue: 20000
stepvalue: 40000
iter_size: 1
type: "RMSProp"
eval_type: "detection"
ap_version: "11point"

我在這里刪除了test的部分，如果需要的話可以對照原來的版本進行修改。

其他

以上就是整個訓練過程的記錄，期間還有一些小 bug 沒記下來，但應該問題不大，網上查查就解決了。

測試過程和訓練過程大同小異，使用 gen.py 那個腳本產生一個test.prototxt就行了

建議編譯時添加caffe-ssd的python接口，因為很多用於推理的腳本都是python寫的，不然需要自己寫。