Caffe VGG训练图片分类任务

准备数据集

train.txt   test.txt
每行一个图片绝对路径+空格+类别标号，eg “绝对path/001.jpg 0”

构建model

VGG主页 http://www.robots.ox.ac.uk/~vgg/research/very_deep/
下载weights（VGG_ILSVRC_16_layers.caffemodel）和网络prototxt（VGG_ILSVRC_16_layers_deploy.prototxt）

VGG_ILSVRC_16_layers_deploy.prototxt 只包含vgg的backbone部分，还需要自己加入data layer和loss layer
除特殊声明，涉及的路径均为相对于caffe_root的相对路径。

VGG_train_val.prototxt

name: "VGG_ILSVRC_16_layers"
layer {
  name: "data"
  type: "ImageData" 
  # type: "AnnotatedData"
  include {
    phase: TRAIN
  }
 transform_param {
    #crop_size: 224
    mean_value: 104
    mean_value: 117
    mean_value: 123
    mirror: true
 }
 image_data_param {
    source: "examples/garbage/trainval.txt"
    batch_size: 8
    new_height: 224
    new_width: 224
  }
  # data_param {
  #   source: "examples/garbage/garbage_trainval_lmdb"
  #   batch_size: 8
  #   backend: LMDB
  # }
  top: "data"
  top: "label"
}
layer {
  name: "data"
  type: "ImageData" 
  # type: "AnnotatedData"
  include {
    phase: TEST
  }
 transform_param {
    #crop_size: 224
    mean_value: 104
    mean_value: 117
    mean_value: 123
    mirror: false
 }
 image_data_param {
    source: "examples/garbage/trainval.txt"
    batch_size: 8
    new_height: 224
    new_width: 224
  }
  # data_param {
  #   source: "examples/garbage/garbage_trainval_lmdb"
  #   batch_size: 8
  #   backend: LMDB
  # }
  top: "data"
  top: "label"
}
###########################################
# VGG_ILSVRC_16_layers_deploy.prototxt
##########################################
layer {
  name: "loss"
  type: "SoftmaxWithLoss"
  bottom: "fc8"
  bottom: "label"
  top: "loss/loss"
}
layer {
  name: "accuracy/top1"
  type: "Accuracy"
  bottom: "fc8"
  bottom: "label"
  top: "accuracy@1"
  include: { phase: TEST }
  accuracy_param {
    top_k: 1
  }
}
layer {
  name: "accuracy/top5"
  type: "Accuracy"
  bottom: "fc8"
  bottom: "label"
  top: "accuracy@5"
  include: { phase: TEST }
  accuracy_param {
    top_k: 5
  }
}

注意，原VGG用于imagenet分类任务，有1000类别，所有fc8的num_output为1000，因此需要将fc8的num_output改成自己的类别数。同时修改fc8的name，比如改成“fc8_new”，这样加载预训练model时便会跳过这一层，否则会因为shape不匹配报错。（当然也可以直接修改VGG_ILSVRC_16_layers.caffemodel，将fc8去掉）

layer {
  name: "fc8_new"
  bottom: "fc7"
  top: "fc8"
  type: "InnerProduct"
  inner_product_param {
    num_output: 40 #1000
    weight_filler {
      type: "xavier"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
  param {
    lr_mult: 1
    decay_mult :1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
}

solver.prototxt

net: "models/garbage/512x512/VGG_train_val.prototxt"
test_iter: 10000
test_interval: 40000
test_initialization: false
display: 200
base_lr: 0.0001
lr_policy: "step"
stepsize: 320000
gamma: 0.1
max_iter: 10000000
momentum: 0.9
weight_decay: 0.0005
snapshot: 800000
snapshot_prefix: "models/vgg/vgg"
solver_mode: GPU

train命令

cd [caffe_root]
./build/tools/caffe train \
--solver="models/garbage/VGG_512x512/solver.prototxt" \
--weights="models/VGG_ILSVRC_16_layers.caffemodel" \
--gpu 0,1 2>&1 

test命令

cd [caffe_root]
./build/tools/caffe test \
--model="models/garbage/VGG_512x512/VGG_train_val.prototxt" \
--weights="models/garbage/VGG_512x512/VGG_iter_XX.caffemodel" \
--gpu 0,1 2>&1