1

我正在尝试基于 Imagenet 示例添加一个 dropout 层（请参见下面的代码）。但是，它似乎被忽略了，当我训练模型并收到下面的警告消息时，它没有被打印为网络的一部分。我安装了最新版本的 caffe。我应该怎么做才能正确包含它？

辍学层：

layer {
  name: "drop4"
  type: "Dropout"
  bottom: "fc4"
  top: "fc4"
  dropout_param {
    dropout_ratio: 0.5
  }
}

训练日志：

I0407 21:05:06.809166  5962 caffe.cpp:117] Use CPU.
I0407 21:05:06.809468  5962 caffe.cpp:121] Starting Optimization
I0407 21:05:06.809532  5962 solver.cpp:32] Initializing solver from parameters:
test_iter: 100
test_interval: 500
base_lr: 0.01
display: 100
max_iter: 10000
lr_policy: "fixed"
gamma: 0.0001
power: 0.75
weight_decay: 0.0005
snapshot: 1000
snapshot_prefix: "hdf5_classification/data/train"
solver_mode: CPU
net: "hdf5_classification/cnn_train.prototxt"
solver_type: ADAGRAD
I0407 21:05:06.809566  5962 solver.cpp:70] Creating training net from net file: hdf5_classification/cnn_train.prototxt
E0407 21:05:06.809836  5962 upgrade_proto.cpp:618] Attempting to upgrade input file specified using deprecated V1LayerParameter: hdf5_classification/cnn_train.prototxt
E0407 21:05:06.809968  5962 upgrade_proto.cpp:636] Input NetParameter to be upgraded already specifies 'layer' fields; these will be ignored for the upgrade.
E0407 21:05:06.810035  5962 upgrade_proto.cpp:623] Warning: had one or more problems upgrading V1LayerParameter (see above); continuing anyway.
I0407 21:05:06.810108  5962 net.cpp:257] The NetState phase (0) differed from the phase (1) specified by a rule in layer data
I0407 21:05:06.810132  5962 net.cpp:257] The NetState phase (0) differed from the phase (1) specified by a rule in layer accuracy
I0407 21:05:06.810143  5962 net.cpp:257] The NetState phase (0) differed from the phase (1) specified by a rule in layer pred
I0407 21:05:06.810266  5962 net.cpp:42] Initializing net from parameters:
name: "CDR-CNN"
state {
  phase: TRAIN
}
layer {
  name: "data"
  type: "HDF5Data"
  top: "data"
  top: "label"
  include {
    phase: TRAIN
  }
  hdf5_data_param {
    source: "hdf5_classification/data/train.txt"
    batch_size: 10
  }
}
layer {
  name: "conv1"
  type: "Convolution"
  bottom: "data"
  top: "conv1"
  param {
    lr_mult: 1
  }
  param {
    lr_mult: 2
  }
  convolution_param {
    num_output: 16
    weight_filler {
      type: "gaussian"
      std: 0.01
    }
    bias_filler {
      type: "constant"
    }
    kernel_h: 1
    kernel_w: 3
    stride_h: 1
    stride_w: 1
  }
}
layer {
  name: "relu1"
  type: "ReLU"
  bottom: "conv1"
  top: "conv1"
}
layer {
  name: "pool1"
  type: "Pooling"
  bottom: "conv1"
  top: "pool1"
  pooling_param {
    pool: MAX
    kernel_h: 1
    kernel_w: 2
    stride_h: 1
    stride_w: 2
  }
}
layer {
  name: "conv2"
  type: "Convolution"
  bottom: "pool1"
  top: "conv2"
  param {
    lr_mult: 1
  }
  param {
    lr_mult: 2
  }
  convolution_param {
    num_output: 20
    weight_filler {
      type: "gaussian"
      std: 0.01
    }
    bias_filler {
      type: "constant"
    }
    kernel_h: 1
    kernel_w: 11
    stride_h: 1
    stride_w: 1
  }
}
layer {
  name: "relu2"
  type: "ReLU"
  bottom: "conv2"
  top: "conv2"
}
layer {
  name: "conv3"
  type: "Convolution"
  bottom: "conv2"
  top: "conv3"
  param {
    lr_mult: 1
  }
  param {
    lr_mult: 2
  }
  convolution_param {
    num_output: 120
    weight_filler {
      type: "gaussian"
      std: 0.01
    }
    bias_filler {
      type: "constant"
    }
    kernel_h: 7
    kernel_w: 1
    stride_h: 1
    stride_w: 1
  }
}
layer {
  name: "relu3"
  type: "ReLU"
  bottom: "conv3"
  top: "conv3"
}
layer {
  name: "fc4"
  type: "InnerProduct"
  bottom: "conv3"
  top: "fc4"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  inner_product_param {
    num_output: 84
    weight_filler {
      type: "gaussian"
      std: 0.01
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  name: "relu4"
  type: "ReLU"
  bottom: "fc4"
  top: "fc4"
}
layer {
  name: "fc5"
  type: "InnerProduct"
  bottom: "fc4"
  top: "fc5"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  inner_product_param {
    num_output: 2
    weight_filler {
      type: "gaussian"
      std: 0.01
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  name: "loss"
  type: "SoftmaxWithLoss"
  bottom: "fc5"
  bottom: "label"
  top: "loss"
  include {
    phase: TRAIN
  }
}
I0407 21:05:06.810355  5962 layer_factory.hpp:74] Creating layer data
I0407 21:05:06.811199  5962 net.cpp:84] Creating Layer data
I0407 21:05:06.811228  5962 net.cpp:338] data -> data
I0407 21:05:06.811259  5962 net.cpp:338] data -> label
I0407 21:05:06.811285  5962 net.cpp:113] Setting up data
I0407 21:05:06.811301  5962 hdf5_data_layer.cpp:80] Loading list of HDF5 filenames from: hdf5_classification/data/train.txt
I0407 21:05:06.811357  5962 hdf5_data_layer.cpp:94] Number of HDF5 files: 1
I0407 21:05:06.853078  5962 net.cpp:120] Top shape: 10 14 7 24 (23520)
I0407 21:05:06.853113  5962 net.cpp:120] Top shape: 10 (10)
I0407 21:05:06.853132  5962 layer_factory.hpp:74] Creating layer conv1
I0407 21:05:06.853159  5962 net.cpp:84] Creating Layer conv1
I0407 21:05:06.853175  5962 net.cpp:380] conv1 <- data
I0407 21:05:06.853199  5962 net.cpp:338] conv1 -> conv1
I0407 21:05:06.853221  5962 net.cpp:113] Setting up conv1
I0407 21:05:06.853638  5962 net.cpp:120] Top shape: 10 16 7 22 (24640)
I0407 21:05:06.853665  5962 layer_factory.hpp:74] Creating layer relu1
I0407 21:05:06.853684  5962 net.cpp:84] Creating Layer relu1
I0407 21:05:06.853698  5962 net.cpp:380] relu1 <- conv1
I0407 21:05:06.853713  5962 net.cpp:327] relu1 -> conv1 (in-place)
I0407 21:05:06.853727  5962 net.cpp:113] Setting up relu1
I0407 21:05:06.853744  5962 net.cpp:120] Top shape: 10 16 7 22 (24640)
I0407 21:05:06.853757  5962 layer_factory.hpp:74] Creating layer pool1
I0407 21:05:06.853772  5962 net.cpp:84] Creating Layer pool1
I0407 21:05:06.853785  5962 net.cpp:380] pool1 <- conv1
I0407 21:05:06.853799  5962 net.cpp:338] pool1 -> pool1
I0407 21:05:06.853814  5962 net.cpp:113] Setting up pool1
I0407 21:05:06.853837  5962 net.cpp:120] Top shape: 10 16 7 11 (12320)
I0407 21:05:06.853849  5962 layer_factory.hpp:74] Creating layer conv2
I0407 21:05:06.853867  5962 net.cpp:84] Creating Layer conv2
I0407 21:05:06.853878  5962 net.cpp:380] conv2 <- pool1
I0407 21:05:06.853893  5962 net.cpp:338] conv2 -> conv2
I0407 21:05:06.853909  5962 net.cpp:113] Setting up conv2
I0407 21:05:06.854030  5962 net.cpp:120] Top shape: 10 20 7 1 (1400)
I0407 21:05:06.854048  5962 layer_factory.hpp:74] Creating layer relu2
I0407 21:05:06.854063  5962 net.cpp:84] Creating Layer relu2
I0407 21:05:06.854074  5962 net.cpp:380] relu2 <- conv2
I0407 21:05:06.854087  5962 net.cpp:327] relu2 -> conv2 (in-place)
I0407 21:05:06.854100  5962 net.cpp:113] Setting up relu2
I0407 21:05:06.854113  5962 net.cpp:120] Top shape: 10 20 7 1 (1400)
I0407 21:05:06.854125  5962 layer_factory.hpp:74] Creating layer conv3
I0407 21:05:06.854140  5962 net.cpp:84] Creating Layer conv3
I0407 21:05:06.854152  5962 net.cpp:380] conv3 <- conv2
I0407 21:05:06.854166  5962 net.cpp:338] conv3 -> conv3
I0407 21:05:06.854179  5962 net.cpp:113] Setting up conv3
I0407 21:05:06.854748  5962 net.cpp:120] Top shape: 10 120 1 1 (1200)
I0407 21:05:06.854771  5962 layer_factory.hpp:74] Creating layer relu3
I0407 21:05:06.854785  5962 net.cpp:84] Creating Layer relu3
I0407 21:05:06.854797  5962 net.cpp:380] relu3 <- conv3
I0407 21:05:06.854811  5962 net.cpp:327] relu3 -> conv3 (in-place)
I0407 21:05:06.854825  5962 net.cpp:113] Setting up relu3
I0407 21:05:06.854838  5962 net.cpp:120] Top shape: 10 120 1 1 (1200)
I0407 21:05:06.854851  5962 layer_factory.hpp:74] Creating layer fc4
I0407 21:05:06.854871  5962 net.cpp:84] Creating Layer fc4
I0407 21:05:06.854883  5962 net.cpp:380] fc4 <- conv3
I0407 21:05:06.854897  5962 net.cpp:338] fc4 -> fc4
I0407 21:05:06.854912  5962 net.cpp:113] Setting up fc4
I0407 21:05:06.855232  5962 net.cpp:120] Top shape: 10 84 (840)
I0407 21:05:06.855252  5962 layer_factory.hpp:74] Creating layer relu4
I0407 21:05:06.855267  5962 net.cpp:84] Creating Layer relu4
I0407 21:05:06.855278  5962 net.cpp:380] relu4 <- fc4
I0407 21:05:06.855406  5962 net.cpp:327] relu4 -> fc4 (in-place)
I0407 21:05:06.855432  5962 net.cpp:113] Setting up relu4
I0407 21:05:06.855447  5962 net.cpp:120] Top shape: 10 84 (840)
I0407 21:05:06.855458  5962 layer_factory.hpp:74] Creating layer fc5
I0407 21:05:06.855582  5962 net.cpp:84] Creating Layer fc5
I0407 21:05:06.855614  5962 net.cpp:380] fc5 <- fc4
I0407 21:05:06.855631  5962 net.cpp:338] fc5 -> fc5
I0407 21:05:06.855648  5962 net.cpp:113] Setting up fc5
I0407 21:05:06.855674  5962 net.cpp:120] Top shape: 10 2 (20)
I0407 21:05:06.855690  5962 layer_factory.hpp:74] Creating layer loss
I0407 21:05:06.855710  5962 net.cpp:84] Creating Layer loss
I0407 21:05:06.855721  5962 net.cpp:380] loss <- fc5
I0407 21:05:06.855734  5962 net.cpp:380] loss <- label
I0407 21:05:06.855751  5962 net.cpp:338] loss -> loss
I0407 21:05:06.855768  5962 net.cpp:113] Setting up loss
I0407 21:05:06.855785  5962 layer_factory.hpp:74] Creating layer loss
I0407 21:05:06.855813  5962 net.cpp:120] Top shape: (1)
I0407 21:05:06.855825  5962 net.cpp:122]     with loss weight 1
I0407 21:05:06.855854  5962 net.cpp:167] loss needs backward computation.
I0407 21:05:06.855865  5962 net.cpp:167] fc5 needs backward computation.
I0407 21:05:06.855877  5962 net.cpp:167] relu4 needs backward computation.
I0407 21:05:06.855890  5962 net.cpp:167] fc4 needs backward computation.
I0407 21:05:06.855901  5962 net.cpp:167] relu3 needs backward computation.
I0407 21:05:06.855912  5962 net.cpp:167] conv3 needs backward computation.
I0407 21:05:06.855924  5962 net.cpp:167] relu2 needs backward computation.
I0407 21:05:06.855937  5962 net.cpp:167] conv2 needs backward computation.
I0407 21:05:06.855947  5962 net.cpp:167] pool1 needs backward computation.
I0407 21:05:06.855959  5962 net.cpp:167] relu1 needs backward computation.
I0407 21:05:06.855970  5962 net.cpp:167] conv1 needs backward computation.
I0407 21:05:06.855983  5962 net.cpp:169] data does not need backward computation.
I0407 21:05:06.855994  5962 net.cpp:205] This network produces output loss
I0407 21:05:06.856011  5962 net.cpp:447] Collecting Learning Rate and Weight Decay.
I0407 21:05:06.856029  5962 net.cpp:217] Network initialization done.
I0407 21:05:06.856041  5962 net.cpp:218] Memory required for data: 368124
4

1 回答 1

2

我遇到了完全相同的问题，偶然发现了这个问题：

“尝试升级使用已弃用的转换参数指定的输入文件”是什么意思？

据我所知，在 caffe 尝试升级你的层架构时，caffe 已经找到了一些用新的 protobuf 语法编写的层。因此，它会跳过升级旧语法，从而忽略 Dropout 层。但我不是 100% 确定这一点，因为这只是我的假设，在简要查看 caffe 代码时。

对我来说，解决方案是将其更改如下：

layers {
  name: "drop4"
  type: DROPOUT
  bottom: "fc4"
  top: "fc4"
  dropout_param {
    dropout_ratio: 0.5
  }
}

希望它有效 =)

于 2015-08-25T10:12:15.157 回答