在FCN训练时,loss一直保持不变。我的分类数不是21类,是2类,请问该怎么初始化Deconvolution的参数?

在训练FCN的时候,使用的二值化的cityscapes数据集,我的分类数目为2.我的train.prototxt如下所示:
 
name: "fcn8snet"
layer {
  name: "data"
  type: "ImageData"
  top: "data"
  top: "fake-dlabel"
  include {
    phase: TRAIN
  }
  transform_param {
    mean_file: "fcn8s_cityscapes/fcn_mean.binaryproto"
    #scale: 0.00390625
  }
  image_data_param {
    source: "fcn8s_cityscapes/data/train_img.txt"
    batch_size: 1
    root_folder: "fcn8s_cityscapes/data/train/image/"
  }
}
layer {
  name: "label"
  type: "ImageData"
  top: "label"
  top: "fake_llabel"
  include {
    phase: TRAIN
  }
  image_data_param {
    source: "fcn8s_cityscapes/data/train_label.txt"
    batch_size: 1
    root_folder: "fcn8s_cityscapes/data/train/label/"
    is_color: false
  }
}
layer {
  name: "conv1_1"
  type: "Convolution"
  bottom: "data"
  top: "conv1_1"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 64
    pad: 100
    kernel_size: 3
    stride: 1
  }
}
layer {
  name: "relu1_1"
  type: "ReLU"
  bottom: "conv1_1"
  top: "conv1_1"
}
layer {
  name: "conv1_2"
  type: "Convolution"
  bottom: "conv1_1"
  top: "conv1_2"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 64
    pad: 1
    kernel_size: 3
    stride: 1
  }
}
layer {
  name: "relu1_2"
  type: "ReLU"
  bottom: "conv1_2"
  top: "conv1_2"
}
layer {
  name: "pool1"
  type: "Pooling"
  bottom: "conv1_2"
  top: "pool1"
  pooling_param {
    pool: MAX
    kernel_size: 2
    stride: 2
  }
}
layer {
  name: "conv2_1"
  type: "Convolution"
  bottom: "pool1"
  top: "conv2_1"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 128
    pad: 1
    kernel_size: 3
    stride: 1
  }
}
layer {
  name: "relu2_1"
  type: "ReLU"
  bottom: "conv2_1"
  top: "conv2_1"
}
layer {
  name: "conv2_2"
  type: "Convolution"
  bottom: "conv2_1"
  top: "conv2_2"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 128
    pad: 1
    kernel_size: 3
    stride: 1
  }
}
layer {
  name: "relu2_2"
  type: "ReLU"
  bottom: "conv2_2"
  top: "conv2_2"
}
layer {
  name: "pool2"
  type: "Pooling"
  bottom: "conv2_2"
  top: "pool2"
  pooling_param {
    pool: MAX
    kernel_size: 2
    stride: 2
  }
}
layer {
  name: "conv3_1"
  type: "Convolution"
  bottom: "pool2"
  top: "conv3_1"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 256
    pad: 1
    kernel_size: 3
    stride: 1
  }
}
layer {
  name: "relu3_1"
  type: "ReLU"
  bottom: "conv3_1"
  top: "conv3_1"
}
layer {
  name: "conv3_2"
  type: "Convolution"
  bottom: "conv3_1"
  top: "conv3_2"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 256
    pad: 1
    kernel_size: 3
    stride: 1
  }
}
layer {
  name: "relu3_2"
  type: "ReLU"
  bottom: "conv3_2"
  top: "conv3_2"
}
layer {
  name: "conv3_3"
  type: "Convolution"
  bottom: "conv3_2"
  top: "conv3_3"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 256
    pad: 1
    kernel_size: 3
    stride: 1
  }
}
layer {
  name: "relu3_3"
  type: "ReLU"
  bottom: "conv3_3"
  top: "conv3_3"
}
layer {
  name: "pool3"
  type: "Pooling"
  bottom: "conv3_3"
  top: "pool3"
  pooling_param {
    pool: MAX
    kernel_size: 2
    stride: 2
  }
}
layer {
  name: "conv4_1"
  type: "Convolution"
  bottom: "pool3"
  top: "conv4_1"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 512
    pad: 1
    kernel_size: 3
    stride: 1
  }
}
layer {
  name: "relu4_1"
  type: "ReLU"
  bottom: "conv4_1"
  top: "conv4_1"
}
layer {
  name: "conv4_2"
  type: "Convolution"
  bottom: "conv4_1"
  top: "conv4_2"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 512
    pad: 1
    kernel_size: 3
    stride: 1
  }
}
layer {
  name: "relu4_2"
  type: "ReLU"
  bottom: "conv4_2"
  top: "conv4_2"
}
layer {
  name: "conv4_3"
  type: "Convolution"
  bottom: "conv4_2"
  top: "conv4_3"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 512
    pad: 1
    kernel_size: 3
    stride: 1
  }
}
layer {
  name: "relu4_3"
  type: "ReLU"
  bottom: "conv4_3"
  top: "conv4_3"
}
layer {
  name: "pool4"
  type: "Pooling"
  bottom: "conv4_3"
  top: "pool4"
  pooling_param {
    pool: MAX
    kernel_size: 2
    stride: 2
  }
}
layer {
  name: "conv5_1"
  type: "Convolution"
  bottom: "pool4"
  top: "conv5_1"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 512
    pad: 1
    kernel_size: 3
    stride: 1
  }
}
layer {
  name: "relu5_1"
  type: "ReLU"
  bottom: "conv5_1"
  top: "conv5_1"
}
layer {
  name: "conv5_2"
  type: "Convolution"
  bottom: "conv5_1"
  top: "conv5_2"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 512
    pad: 1
    kernel_size: 3
    stride: 1
  }
}
layer {
  name: "relu5_2"
  type: "ReLU"
  bottom: "conv5_2"
  top: "conv5_2"
}
layer {
  name: "conv5_3"
  type: "Convolution"
  bottom: "conv5_2"
  top: "conv5_3"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 512
    pad: 1
    kernel_size: 3
    stride: 1
  }
}
layer {
  name: "relu5_3"
  type: "ReLU"
  bottom: "conv5_3"
  top: "conv5_3"
}
layer {
  name: "pool5"
  type: "Pooling"
  bottom: "conv5_3"
  top: "pool5"
  pooling_param {
    pool: MAX
    kernel_size: 2
    stride: 2
  }
}
layer {
  name: "fc6"
  type: "Convolution"
  bottom: "pool5"
  top: "fc6"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 4096
    pad: 0
    kernel_size: 7
    stride: 1
  }
}
layer {
  name: "relu6"
  type: "ReLU"
  bottom: "fc6"
  top: "fc6"
}
layer {
  name: "drop6"
  type: "Dropout"
  bottom: "fc6"
  top: "fc6"
  dropout_param {
    dropout_ratio: 0.5
  }
}
layer {
  name: "fc7"
  type: "Convolution"
  bottom: "fc6"
  top: "fc7"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 4096
    pad: 0
    kernel_size: 1
    stride: 1
  }
}
layer {
  name: "relu7"
  type: "ReLU"
  bottom: "fc7"
  top: "fc7"
}
layer {
  name: "drop7"
  type: "Dropout"
  bottom: "fc7"
  top: "fc7"
  dropout_param {
    dropout_ratio: 0.5
  }
}
layer {
  name: "score_fr_cityscapes"
  type: "Convolution"
  bottom: "fc7"
  top: "score_fr"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 2
    pad: 0
    kernel_size: 1
    weight_filler { type: "gaussian" std: 0.01 }
    bias_filler { type: "constant" value: 0 }
  }
}
layer {
  name: "upscore2_cityscapes"
  type: "Deconvolution"
  bottom: "score_fr"
  top: "upscore2"
  param {
    lr_mult: 0
  }
  convolution_param {
    num_output: 2
    bias_term: false
    kernel_size: 4
    stride: 2
    weight_filler { type: "gaussian" std: 0.01 }
    bias_filler { type: "constant" value: 0 }
  }
}
layer {
  name: "score_pool4_cityscapes"
  type: "Convolution"
  bottom: "pool4"
  top: "score_pool4"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 2
    pad: 0
    kernel_size: 1
    weight_filler { type: "gaussian" std: 0.01 }
    bias_filler { type: "constant" value: 0 }
  }
}
layer {
  name: "score_pool4c"
  type: "Crop"
  bottom: "score_pool4"
  bottom: "upscore2"
  top: "score_pool4c"
  crop_param {
    axis: 2
    offset: 5
  }
}
layer {
  name: "fuse_pool4"
  type: "Eltwise"
  bottom: "upscore2"
  bottom: "score_pool4c"
  top: "fuse_pool4"
  eltwise_param {
    operation: SUM
  }
}
layer {
  name: "upscore_pool4_cityscapes"
  type: "Deconvolution"
  bottom: "fuse_pool4"
  top: "upscore_pool4"
  param {
    lr_mult: 0
  }
  convolution_param {
    num_output: 2
    bias_term: false
    kernel_size: 4
    stride: 2
    weight_filler { type: "gaussian" std: 0.01 }
    bias_filler { type: "constant" value: 0 }
  }
}
layer {
  name: "score_pool3_cityscapes"
  type: "Convolution"
  bottom: "pool3"
  top: "score_pool3"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 2
    pad: 0
    kernel_size: 1
    weight_filler { type: "gaussian" std: 0.01 }
    bias_filler { type: "constant" value: 0 }
  }
}
layer {
  name: "score_pool3c"
  type: "Crop"
  bottom: "score_pool3"
  bottom: "upscore_pool4"
  top: "score_pool3c"
  crop_param {
    axis: 2
    offset: 9
  }
}
layer {
  name: "fuse_pool3"
  type: "Eltwise"
  bottom: "upscore_pool4"
  bottom: "score_pool3c"
  top: "fuse_pool3"
  eltwise_param {
    operation: SUM
  }
}
layer {
  name: "upscore8_cityscapes"
  type: "Deconvolution"
  bottom: "fuse_pool3"
  top: "upscore8"
  param {
    lr_mult: 0
  }
  convolution_param {
    num_output: 2
    bias_term: false
    kernel_size: 16
    stride: 8
    weight_filler { type: "gaussian" std: 0.01 }
    bias_filler { type: "constant" value: 0 }
  }
}
layer {
  name: "score"
  type: "Crop"
  bottom: "upscore8"
  bottom: "data"
  top: "score"
  crop_param {
    axis: 2
    offset: 31
  }
}
layer {
  name: "loss"
  type: "SoftmaxWithLoss"
  bottom: "score"
  bottom: "label"
  top: "loss"
  loss_param {
    ignore_label: 255
    #normalize: false
  }
}
 
可是我的loss一直稳定在0.69附近:

11111.jpg

有听到其他链接里面说到,反卷积层参数可以通过fine-tune初始化的方式解决
http://www.caffecn.cn/?/question/612
我想请问,在修改分类数的前提下,反卷积层的参数怎么初始化?怎么才能让FCN正常训练起来?
 
拜托大家!
 
 
已邀请:

jiongnima

赞同来自: caffe

本问题已经解决,用官方提供的python_layer,配置训练完成,得到正确结果。

jiongnima

赞同来自: loveu

训练不收敛的问题解决了,原因是反卷积层初始化的问题。
解决方案是使用作者提供的数据层(python_layer),并且在反卷积层名中保留"up"字样。
详见:http://blog.csdn.net/jiongnima ... 49326

要回复问题请先登录注册