faster rcnn 当中stage1_rpn_train.pt的解读

#stage 1训练RPN时用的网络结构  
name: "ZF"
layer {
  name: 'input-data'
  type: 'Python'
  top: 'data'
  top: 'im_info'
  top: 'gt_boxes'
  python_param {
    module: 'roi_data_layer.layer'#对应lib/roi_data_layer/layer.py为训练RPN时为网络输入roi,此时为gt box  
    layer: 'RoIDataLayer'
    param_str: "'num_classes': 21"
  }
}

#========= conv1-conv5 ============

layer {
    name: "conv1"
    type: "Convolution"
    bottom: "data"
    top: "conv1"
    param { lr_mult: 1.0 }
    param { lr_mult: 2.0 }
    convolution_param {
        num_output: 96
        kernel_size: 7
        pad: 3
        stride: 2
    }
}
layer {
    name: "relu1"
    type: "ReLU"
    bottom: "conv1"
    top: "conv1"
}
layer {
    name: "norm1"
    type: "LRN"
    bottom: "conv1"
    top: "norm1"
    lrn_param {
        local_size: 3
        alpha: 0.00005
        beta: 0.75
        norm_region: WITHIN_CHANNEL
    engine: CAFFE
    }
}
layer {
    name: "pool1"
    type: "Pooling"
    bottom: "norm1"
    top: "pool1"
    pooling_param {
        kernel_size: 3
        stride: 2
        pad: 1
        pool: MAX
    }
}
layer {
    name: "conv2"
    type: "Convolution"
    bottom: "pool1"
    top: "conv2"
    param { lr_mult: 1.0 }
    param { lr_mult: 2.0 }
    convolution_param {
        num_output: 256
        kernel_size: 5
        pad: 2
        stride: 2
    }
}
layer {
    name: "relu2"
    type: "ReLU"
    bottom: "conv2"
    top: "conv2"
}
layer {
    name: "norm2"
    type: "LRN"
    bottom: "conv2"
    top: "norm2"
    lrn_param {
        local_size: 3
        alpha: 0.00005
        beta: 0.75
        norm_region: WITHIN_CHANNEL
    engine: CAFFE
    }
}
layer {
    name: "pool2"
    type: "Pooling"
    bottom: "norm2"
    top: "pool2"
    pooling_param {
        kernel_size: 3
        stride: 2
        pad: 1
        pool: MAX
    }
}
layer {
    name: "conv3"
    type: "Convolution"
    bottom: "pool2"
    top: "conv3"
    param { lr_mult: 1.0 }
    param { lr_mult: 2.0 }
    convolution_param {
        num_output: 384
        kernel_size: 3
        pad: 1
        stride: 1
    }
}
layer {
    name: "relu3"
    type: "ReLU"
    bottom: "conv3"
    top: "conv3"
}
layer {
    name: "conv4"
    type: "Convolution"
    bottom: "conv3"
    top: "conv4"
    param { lr_mult: 1.0 }
    param { lr_mult: 2.0 }
    convolution_param {
        num_output: 384
        kernel_size: 3
        pad: 1
        stride: 1
    }
}
layer {
    name: "relu4"
    type: "ReLU"
    bottom: "conv4"
    top: "conv4"
}
layer {
    name: "conv5"
    type: "Convolution"
    bottom: "conv4"
    top: "conv5"
    param { lr_mult: 1.0 }
    param { lr_mult: 2.0 }
    convolution_param {
        num_output: 256
        kernel_size: 3
        pad: 1
        stride: 1
    }
}
layer {
    name: "relu5"
    type: "ReLU"
    bottom: "conv5"
    top: "conv5"
}

#========= RPN ============
#对于前景和背景进行筛选【二分类】

layer {
  name: "rpn_conv1"
  type: "Convolution"
  bottom: "conv5"
  top: "rpn_conv1"
  param { lr_mult: 1.0 }
  param { lr_mult: 2.0 }
  convolution_param {
    num_output: 256
    kernel_size: 3 pad: 1 stride: 1#生成3*3的滑动窗口
    weight_filler { type: "gaussian" std: 0.01 }
    bias_filler { type: "constant" value: 0 }
  }
}
layer {
  name: "rpn_relu1"
  type: "ReLU"
  bottom: "rpn_conv1"
  top: "rpn_conv1"
}
layer {
  name: "rpn_cls_score"
  type: "Convolution"
  bottom: "rpn_conv1"
  top: "rpn_cls_score"
  param { lr_mult: 1.0 }
  param { lr_mult: 2.0 }
  convolution_param {
    num_output: 18   # 2(bg/fg) * 9(anchors)#二分类,因此是2×anchors种类
    kernel_size: 1 pad: 0 stride: 1
    weight_filler { type: "gaussian" std: 0.01 }
    bias_filler { type: "constant" value: 0 }
  }
}
layer {
  name: "rpn_bbox_pred"
  type: "Convolution"
  bottom: "rpn_conv1"
  top: "rpn_bbox_pred"
  param { lr_mult: 1.0 }
  param { lr_mult: 2.0 }
  convolution_param {
    num_output: 36   # 4 * 9(anchors)#回归问题,4个坐标点
    kernel_size: 1 pad: 0 stride: 1
    weight_filler { type: "gaussian" std: 0.01 }
    bias_filler { type: "constant" value: 0 }
  }
}
layer {
   bottom: "rpn_cls_score"
   top: "rpn_cls_score_reshape"
   name: "rpn_cls_score_reshape"
   type: "Reshape"
   reshape_param { shape { dim: 0 dim: 2 dim: -1 dim: 0 } }
}
layer {
  name: 'rpn-data'
  type: 'Python'
  bottom: 'rpn_cls_score'
  bottom: 'gt_boxes'
  bottom: 'im_info'
  bottom: 'data'
  top: 'rpn_labels'
  top: 'rpn_bbox_targets'
  top: 'rpn_bbox_inside_weights'
  top: 'rpn_bbox_outside_weights'
  python_param {
    module: 'rpn.anchor_target_layer'##对应文件lib/rpn/anchor_target_layer.py  
#用于在原图上产生anchor,结合gt box训练rpn做box cls和box reg
    layer: 'AnchorTargetLayer'
    param_str: "'feat_stride': 16"
  }
}
layer {
  name: "rpn_loss_cls"#经过rpn层之后的二分类结果
  type: "SoftmaxWithLoss"
  bottom: "rpn_cls_score_reshape"
  bottom: "rpn_labels"
  propagate_down: 1
  propagate_down: 0
  top: "rpn_cls_loss"
  loss_weight: 1
  loss_param {
    ignore_label: -1
    normalize: true
  }
}
layer {
  name: "rpn_loss_bbox"#经过rpn层之后的回归结果
  type: "SmoothL1Loss"
  bottom: "rpn_bbox_pred"
  bottom: "rpn_bbox_targets"
  bottom: "rpn_bbox_inside_weights"
  bottom: "rpn_bbox_outside_weights"
  top: "rpn_loss_bbox"
  loss_weight: 1
  smooth_l1_loss_param { sigma: 3.0 }
}

#=========RCNN ============
#fast rcnn可以对region proposal进行细分和位置精修
# Dummy layers so that initial parameters are saved into the output net
#在rpn层的时候,会产生9种anchor[框的对称中心点],必须经过pooling层将这9个anchor合并成一个,才能继续输入到全连接层fc6
layer {
  name: "dummy_roi_pool_conv5"
  type: "DummyData"
  top: "dummy_roi_pool_conv5"
  dummy_data_param {
    shape { dim: 1 dim: 9216 }
    data_filler { type: "gaussian" std: 0.01 }
  }
}
layer {
  name: "fc6"
  type: "InnerProduct"
  bottom: "dummy_roi_pool_conv5"
  top: "fc6"
  param { lr_mult: 0 decay_mult: 0 }
  param { lr_mult: 0 decay_mult: 0 }
  inner_product_param {
    num_output: 4096
  }
}
layer {
  name: "relu6"
  type: "ReLU"
  bottom: "fc6"
  top: "fc6"
}
layer {
  name: "fc7"#进行最后的分类任务
  type: "InnerProduct"
  bottom: "fc6"
  top: "fc7"
  param { lr_mult: 0 decay_mult: 0 }
  param { lr_mult: 0 decay_mult: 0 }
  inner_product_param {
    num_output: 4096
  }
}
layer {
  name: "silence_fc7"
  type: "Silence"
  bottom: "fc7"
}
  • 0
    点赞
  • 0
    收藏
    觉得还不错? 一键收藏
  • 0
    评论
评论
添加红包

请填写红包祝福语或标题

红包个数最小为10个

红包金额最低5元

当前余额3.43前往充值 >
需支付:10.00
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明:

1.余额是钱包充值的虚拟货币,按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载,可以购买VIP、付费专栏及课程。

余额充值