检查失败：如何在深层使用hdf5数据层？

Question

我有 data.mat 的火车和标签数据。（我有 200 个训练数据和 6000 个特征，标签是 (-1, +1)，保存在 data.mat 中。

我正在尝试使用 hdf5 和运行 Caffe 转换我的数据（训练和测试）：

load input.mat
hdf5write('my_data.h5', '/new_train_x', single( permute(reshape(new_train_x,[200, 6000, 1, 1]),[4:-1:1] ) ));
hdf5write('my_data.h5', '/label_train', single( permute(reshape(label_train,[200, 1, 1, 1]), [4:-1:1] ) ) , 'WriteMode', 'append' );
hdf5write('my_data_test.h5', '/test_x', single( permute(reshape(test_x,[77, 6000, 1, 1]),[4:-1:1] ) ));
hdf5write('my_data_test.h5', '/label_test', single( permute(reshape(label_test,[77, 1, 1, 1]), [4:-1:1] ) ) , 'WriteMode', 'append' );

（请参阅关于在 Matlab 中将 mat 文件转换为 hdf5）。

我的train_val.prototxt是：

  layer {
  type: "HDF5Data"
  name: "data"
  top: "new_train_x"     # note: same name as in HDF5
  top: "label_train"     # 
  hdf5_data_param {
    source: "file.txt"
    batch_size: 20
  }
  include { phase: TRAIN }
}
layer {
  type: "HDF5Data"
  name: "data"
  top: "test_x"     # note: same name as in HDF5
  top: "label_test"     # 
  hdf5_data_param {
    source: "file_test.txt"
    batch_size: 20
  }
  include { phase:TEST }
}

layer {
  name: "ip1"
  type: "InnerProduct"
  bottom: "new_train_x"
  top: "ip1"
  param {
    lr_mult: 1
  }
  param {
    lr_mult: 2
  }
  inner_product_param {
    num_output: 30
    weight_filler {
      type: "gaussian" # initialize the filters from a Gaussian
      std: 0.01 
    }
    bias_filler {
      type: "constant"
    }
  }
}
layer {
  name: "tanh1"
  type: "TanH"
  bottom: "ip1"
  top: "tanh1"
}

layer {
  name: "ip2"
  type: "InnerProduct"
  bottom: "tanh1"
  top: "ip2"
  param {
    lr_mult: 1
  }
  param {
    lr_mult: 2
  }
  inner_product_param {
    num_output: 1
    weight_filler {
      type: "gaussian" # initialize the filters from a Gaussian
      std: 0.01 
    }
    bias_filler {
      type: "constant"
    }
  }
}
layer {
  name: "loss"
  type: "TanH"
  bottom: "ip2"
  bottom: "label_train"
  top: "loss"
}

但是我有一个问题。看来，它无法读取我的输入数据。

I1227 10:27:21.880826  7186 layer_factory.hpp:76] Creating layer data
I1227 10:27:21.880851  7186 net.cpp:110] Creating Layer data
I1227 10:27:21.880866  7186 net.cpp:433] data -> new_train_x
I1227 10:27:21.880893  7186 net.cpp:433] data -> label_train
I1227 10:27:21.880915  7186 hdf5_data_layer.cpp:81] Loading list of HDF5 filenames from: file.txt
I1227 10:27:21.880965  7186 hdf5_data_layer.cpp:95] Number of HDF5 files: 1
I1227 10:27:21.962596  7186 net.cpp:155] Setting up data
I1227 10:27:21.962702  7186 net.cpp:163] Top shape: 20 6000 1 1 (120000)
I1227 10:27:21.962738  7186 net.cpp:163] Top shape: 20 1 1 1 (20)
I1227 10:27:21.962772  7186 layer_factory.hpp:76] Creating layer ip1
I1227 10:27:21.962838  7186 net.cpp:110] Creating Layer ip1
I1227 10:27:21.962873  7186 net.cpp:477] ip1 <- new_train_x
I1227 10:27:21.962918  7186 net.cpp:433] ip1 -> ip1
I1227 10:27:21.979375  7186 net.cpp:155] Setting up ip1
I1227 10:27:21.979434  7186 net.cpp:163] Top shape: 20 30 (600)
I1227 10:27:21.979478  7186 layer_factory.hpp:76] Creating layer tanh1
I1227 10:27:21.979529  7186 net.cpp:110] Creating Layer tanh1
I1227 10:27:21.979557  7186 net.cpp:477] tanh1 <- ip1
I1227 10:27:21.979583  7186 net.cpp:433] tanh1 -> tanh1
I1227 10:27:21.979620  7186 net.cpp:155] Setting up tanh1
I1227 10:27:21.979650  7186 net.cpp:163] Top shape: 20 30 (600)
I1227 10:27:21.979670  7186 layer_factory.hpp:76] Creating layer ip2
I1227 10:27:21.979696  7186 net.cpp:110] Creating Layer ip2
I1227 10:27:21.979720  7186 net.cpp:477] ip2 <- tanh1
I1227 10:27:21.979746  7186 net.cpp:433] ip2 -> ip2
I1227 10:27:21.979796  7186 net.cpp:155] Setting up ip2
I1227 10:27:21.979825  7186 net.cpp:163] Top shape: 20 1 (20)
I1227 10:27:21.979854  7186 layer_factory.hpp:76] Creating layer loss
I1227 10:27:21.979881  7186 net.cpp:110] Creating Layer loss
I1227 10:27:21.979909  7186 net.cpp:477] loss <- ip2
I1227 10:27:21.979931  7186 net.cpp:477] loss <- label_train
I1227 10:27:21.979962  7186 net.cpp:433] loss -> loss
F1227 10:27:21.980006  7186 layer.hpp:374] Check failed: ExactNumBottomBlobs() == bottom.size() (1 vs. 2) TanH Layer takes 1 bottom blob(s) as input.
*** Check failure stack trace: ***
    @     0x7f44cbc68ea4  (unknown)
    @     0x7f44cbc68deb  (unknown)
    @     0x7f44cbc687bf  (unknown)
    @     0x7f44cbc6ba35  (unknown)
    @     0x7f44cbfd0ba8  caffe::Layer<>::CheckBlobCounts()
    @     0x7f44cbfed9da  caffe::Net<>::Init()
    @     0x7f44cbfef108  caffe::Net<>::Net()
    @     0x7f44cc03f71a  caffe::Solver<>::InitTrainNet()
    @     0x7f44cc040a51  caffe::Solver<>::Init()
    @     0x7f44cc040db9  caffe::Solver<>::Solver()
    @           0x41222d  caffe::GetSolver<>()
    @           0x408ed9  train()
    @           0x406741  main
    @     0x7f44ca997a40  (unknown)
    @           0x406f69  _start
    @              (nil)  (unknown)
Aborted (core dumped)

现在，如果我像这样更改损失层：

layer {
  name: "loss"
  type: "TanH"
  bottom: "ip2"
  top: "loss"
}

我有这个问题：

F1227 10:53:17.884419  9102 insert_splits.cpp:35] Unknown bottom blob 'new_train_x' (layer 'ip1', bottom index 0)
*** Check failure stack trace: ***
    @     0x7f502ab5dea4  (unknown)
    @     0x7f502ab5ddeb  (unknown)
    @     0x7f502ab5d7bf  (unknown)
    @     0x7f502ab60a35  (unknown)
    @     0x7f502af1d75b  caffe::InsertSplits()
    @     0x7f502aee19e9  caffe::Net<>::Init()
    @     0x7f502aee4108  caffe::Net<>::Net()
    @     0x7f502af35172  caffe::Solver<>::InitTestNets()
    @     0x7f502af35abd  caffe::Solver<>::Init()
    @     0x7f502af35db9  caffe::Solver<>::Solver()
    @           0x41222d  caffe::GetSolver<>()
    @           0x408ed9  train()
    @           0x406741  main
    @     0x7f502988ca40  (unknown)
    @           0x406f69  _start
    @              (nil)  (unknown)
Aborted (core dumped)

非常感谢！！！！如有任何建议，我们将不胜感激！

Answer 1

您的数据层仅为 phase: TRAIN 定义我相信当 caffe 尝试构建测试时间网络（即 phase: TEST 网络）时会发生错误。
您应该有一个带有测试数据的附加层：

layer {
  type: "HDF5Data"
  name: "data"
  top: "new_train_x"     # note: same name as in HDF5
  top: "label_train"     # 
  hdf5_data_param {
    source: "test_file.txt"
    batch_size: 20
  }
  include { phase: TEST } # do not forget TEST phase
}

顺便说一句，如果你不想在训练过程中测试你的网络，你可以关闭这个选项。有关详细信息，请参阅。

更新：
请原谅我的直言不讳，但你把事情搞得一团糟

"TanH" 不是损失层 - 它是 neuron/activation 层。它用作应用于线性层 (conv/inner-product) 的非线性。因此，它接受单个输入（底部 blob）并输出单个 blob（顶部）。
损失层计算标量损失值，通常需要两个输入：预测和地面实况进行比较。
您确实更改了您的网络并为 TEST 阶段添加了一个 "HDF5Data" 层，但是该层输出了 top: "test_x"，您的网络中没有任何层需要 bottom: "test_x" 你只有层期望 "new_train_x"... 同样适用于 "label_text".

我建议您使用更通用的名称（例如 x 和 label）重新编写 hdf5 文件，以便训练和测试。只需使用不同的文件名来区分它们。这样你的网络在两个阶段都与 "x" 和 "label" 一起工作，并且只根据阶段加载适当的数据集。

检查失败：如何在深层使用hdf5数据层？

Check fail: how to use hdf5 data layer in deep layer?

hdf5

neural-network

deep-learning

caffe