CNN (Caffe) 一维输入的输入层尺寸

如何解决CNN (Caffe) 一维输入的输入层尺寸

我正在尝试在 Caffe 中训练一个简单的 CNN，用于对分成 500 个元素的向量的时间序列数据进行分类。我首先在 Google Colab 中执行此操作：

!apt install caffe-cpu

关注

import caffe

为此，我将数据保存到 hdf5 文件中并定义了三个 .prototxt 文件。训练 sg_train.prototxt 中定义的模型：

layer {
  name: "data"
  type: "HDF5Data"
  top: "data"
  top: "label"
  hdf5_data_param {
    source: "/content/gdrive/MyDrive/data/train2_h5_list.txt"
    batch_size: 32
  }
  include { 
    phase:TRAIN 
  }
}

layer {
  name: "data"
  type: "HDF5Data"
  top: "data"
  top: "label"
  hdf5_data_param {
    source: "/content/gdrive/MyDrive/data/test2_h5_list.txt"
    batch_size: 32
  }
  include { 
    phase:TEST 
  }
}

layer {
  name: "conv1" 
  type: "Convolution" 
  bottom: "data"
  top: "conv1"
  convolution_param {
    num_output: 4 
    kernel_size: 10
    stride: 1 
    pad: 1
  }
}

layer {
  name: "relu1"
  type: "ReLU"
  bottom: "conv1"
  top: "conv1"  
}

layer {
  name: "conv2" 
  type: "Convolution" 
  bottom: "conv1" 
  top: "conv2"
  convolution_param {
    num_output: 8 
    kernel_size: 10
    stride: 1 
    pad: 1
  }
}

layer {
  name: "relu1"
  type: "ReLU"
  bottom: "conv2"
  top: "conv2"  
}

layer {
  name: "fc1" 
  type: "InnerProduct"
  bottom: "conv2"
  top: "fc1"
  inner_product_param {
    num_output: 1
  }
}

layer {
  name: "loss"
  type: "softmaxWithLoss"
  bottom: "fc1"
  bottom: "label"
}

我跑：

caffe.set_mode_cpu()
solver = caffe.get_solver("sg_caffe_solver.prototxt")
solver.solve()

其中 sg_caffe_solver.prototxt 包含：

# The train/test net protocol buffer deFinition
net: "/content/gdrive/MyDrive/models/sg_train.prototxt"
# test_iter specifies how many forward passes the test should carry out.
test_iter: 1
# Carry out testing every test_interval training iterations.
test_interval: 1000
# The base learning rate,momentum and the weight decay of the network.
base_lr: 0.0001
momentum: 0.001
weight_decay: 0.0005
# The learning rate policy
lr_policy: "inv"
gamma: 0.0001
power: 0.75
# display every 100 iterations
display: 1000
# The maximum number of iterations
max_iter: 5000
# snapshot intermediate results
snapshot: 5000
snapshot_prefix: "sg_"
# solver mode: cpu or GPU
solver_mode: cpu # GPU

重要的是，如果我在 sg_train.prototxt 中的 kernel_h 层的定义中使用 kernel_w 和 conv，会话崩溃并且会话日志包含警告 F0209 12:13:53.816366 2919 base_conv_layer.cpp:27] Check Failed: num_spatial_axes_ == 2 (0 vs. 2) kernel_h & kernel_w can only be used for 2D convolution. 否则，在创建 solver.solve() 和 sg__iter_5000.caffemodel 快照的情况下，sg__iter_5000.solverstate 执行时不会出现明显错误（尽管有时内核会重新启动）。

当我后来尝试通过首先使用 net = caffe.Net("sg_deploy.prototxt","sg__iter_5000.caffemodel",caffe.TEST) 实例化 cnn 来测试训练模型时，我遇到了各种错误和内核崩溃，我不明白。 sg_deploy.prototxt 的内容与 sg_train.prototxt 的不同之处在于 HDF5 层被替换为 Input 层：

layer {
  type: "Input"
  name: "data"
  top: "data"
  input_param { shape: { dim: 1 dim: 1 dim: 1 dim: 500 } }
}

然后将 softmaxWithLoss 层替换为 softmax：

layer {
  name: "prob1"
  type: "softmax"
  bottom: "fc1"
  top: "prob1"
}

当我尝试初始化 net 时，内核崩溃了，我在日志中收到以下警告：

F0209 12:40:08.371551 3281 blob.cpp:32] Check Failed: shape[i] >= 0 (-6 vs. 0)
I0209 12:40:08.371512 3281 net.cpp:380] conv1 -> conv1
I0209 12:40:08.371505 3281 net.cpp:406] conv1 <- data

我尝试将 kernel_size = 10 替换为 kernel_w = 10 和 kernel_h = 1。之后初始化 net 也会导致崩溃，但有不同的日志警告：

F0209 12:43:09.321560 3410 net.cpp:757] Cannot copy param 0 weights from layer 'conv1'; shape mismatch. Source param shape is 4 500 (2000); target param shape is 4 1 1 10 (40). To learn this layer's parameters from scratch rather than copying from a saved net,rename the layer.

I0209 12:43:09.310168 3410 net.cpp:255] Network initialization done.

I0209 12:43:09.310160 3410 net.cpp:242] This network produces output prob1

此时我不知道如何继续。