Caffe

caffe-在新样本上使用训练好的模型

用训练好的模型给测试样本打分

经过训练有了模型：my_lr_iter_10000.caffemodel，将新的样本数据的LMDB文件 mnist_newTEST_lmdb 的路径在TEST网络的DataLayer层中指明，将含有标签的新样本文件dataLayer的source：

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
layer{
    name: "mnist"
    type: "Data"  
    top: "data"
    top: "label"
    include {
        phase: TEST  # 该层应用与TEST
    }
    transform_param{  
        scale: 0.0039063
    }
    data_param{   
        source: "/media/junhui/DATA/caffe_workspace/my_linearReggresion/mnist_newTEST_lmdb"
        batch_size: 100  # 每一批 100个样本
        backend: LMDB
    }
}

使用一下命令执行：

1
2
3
4
~/caffe-master/build/tools/caffe test \
-model /media/junhui/DATA/caffe_workspace/my_linearReggresion/mylr.prototxt \
-weights /media/junhui/DATA/caffe_workspace/my_linearReggresion/my_lr_iter_10000.caffemodel \
-iterations 2

test 表示使用TEST网络【只进行前行传播】打分，-model 指明模型结构参数，-weights 指明训练好的模型权值，-iterations 指明对多少样本进行inference，2*100 个样本。

其中caffe实际是可执行文件caffe.bin。

这里是日志文件：

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
I0617 13:28:09.083691  4697 caffe.cpp:275] Use CPU.
I0617 13:28:09.308647  4697 net.cpp:296] The NetState phase (1) differed from the phase (0) specified by a rule in layer mnist
I0617 13:28:09.308768  4697 net.cpp:53] Initializing net from parameters: 
name: "lrNet"
state {
  phase: TEST
  level: 0
  stage: ""
}
layer {
  name: "mnist"
  type: "Data"
  top: "data"
  top: "label"
  include {
    phase: TEST
  }
  transform_param {
    scale: 0.0039063
  }
  data_param {
    source: "/media/junhui/DATA/caffe_workspace/my_linearReggresion/mnist_newtest_lmdb"
    batch_size: 100
    backend: LMDB
  }
}
layer {
  name: "ip"
  type: "InnerProduct"
  bottom: "data"
  top: "ip"
  param {
    lr_mult: 1
  }
  param {
    lr_mult: 2
  }
  inner_product_param {
    num_output: 10
    weight_filler {
      type: "xavier"
    }
    bias_filler {
      type: "constant"
    }
  }
}
layer {
  name: "accuracy"
  type: "Accuracy"
  bottom: "ip"
  bottom: "label"
  top: "accuracy"
  include {
    phase: TEST
  }
}
layer {
  name: "loss"
  type: "SoftmaxWithLoss"
  bottom: "ip"
  bottom: "label"
  top: "loss"
}
I0617 13:28:09.308953  4697 layer_factory.hpp:77] Creating layer mnist
I0617 13:28:09.309533  4697 db_lmdb.cpp:35] Opened lmdb /media/junhui/DATA/caffe_workspace/my_linearReggresion/mnist_newtest_lmdb
# 略
I0617 13:14:44.887392  3080 net.cpp:200] loss needs backward computation.
I0617 13:14:44.887401  3080 net.cpp:202] accuracy does not need backward computation.
I0617 13:14:44.887406  3080 net.cpp:200] ip_ip_0_split needs backward computation.
I0617 13:14:44.887411  3080 net.cpp:200] ip needs backward computation.
I0617 13:14:44.887415  3080 net.cpp:202] label_mnist_1_split does not need backward computation.
I0617 13:14:44.887420  3080 net.cpp:202] mnist does not need backward computation.
I0617 13:14:44.887424  3080 net.cpp:244] This network produces output accuracy
I0617 13:14:44.887430  3080 net.cpp:244] This network produces output loss
I0617 13:14:44.887439  3080 net.cpp:257] Network initialization done.
I0617 13:14:44.888084  3080 caffe.cpp:281] Running for 2 iterations.
# 结果：
I0617 13:14:44.888623  3080 caffe.cpp:304] Batch 0, accuracy = 0.94
I0617 13:14:44.888634  3080 caffe.cpp:304] Batch 0, loss = 0.21204
I0617 13:14:44.889169  3080 caffe.cpp:304] Batch 1, accuracy = 0.96
I0617 13:14:44.889178  3080 caffe.cpp:304] Batch 1, loss = 0.199202
I0617 13:14:44.889200  3080 caffe.cpp:309] Loss: 0.205621
I0617 13:14:44.889210  3080 caffe.cpp:321] accuracy = 0.95
I0617 13:14:44.889219  3080 caffe.cpp:321] loss = 0.205621 (* 1 = 0.205621 loss)

由Log发现，在新样本上使用TEST网络，依然进行了反向传播？因为在TEST网络中要求loss，与accuracy一样是衡量模型在样本上的性能的指标，accuracy越大，loss越小，新能相对越好。而这层需要反向传播计算获得。从尾反向传播到头，所以Log就理所当然了。

使用不带标签的新样本

测试猜想

猜测这种情况下的TEST网络结构只有数据层和全连接层（因为上述网络只有这两个层与label数据无关），最终结果是全连接层的输出。只有两层，进行前向计算（对一批测试，每一批只有一个元素）：

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
...
I0617 13:48:08.206188  5203 caffe.cpp:304] Batch 0, ip = 0.483029
I0617 13:48:08.206202  5203 caffe.cpp:304] Batch 0, ip = -4.20101
I0617 13:48:08.206208  5203 caffe.cpp:304] Batch 0, ip = 0.95857
I0617 13:48:08.206213  5203 caffe.cpp:304] Batch 0, ip = 4.77686
I0617 13:48:08.206219  5203 caffe.cpp:304] Batch 0, ip = -5.86976
I0617 13:48:08.206223  5203 caffe.cpp:304] Batch 0, ip = 5.99296
I0617 13:48:08.206228  5203 caffe.cpp:304] Batch 0, ip = -1.1105
I0617 13:48:08.206233  5203 caffe.cpp:304] Batch 0, ip = 0.394778
I0617 13:48:08.206238  5203 caffe.cpp:304] Batch 0, ip = -0.198426
I0617 13:48:08.206243  5203 caffe.cpp:304] Batch 0, ip = -0.884502
I0617 13:48:08.206248  5203 caffe.cpp:309] Loss: 0
I0617 13:48:08.206256  5203 caffe.cpp:321] ip = 0.483029
I0617 13:48:08.206264  5203 caffe.cpp:321] ip = -4.20101
I0617 13:48:08.206269  5203 caffe.cpp:321] ip = 0.95857
I0617 13:48:08.206272  5203 caffe.cpp:321] ip = 4.77686
I0617 13:48:08.206277  5203 caffe.cpp:321] ip = -5.86976
I0617 13:48:08.206282  5203 caffe.cpp:321] ip = 5.99296
I0617 13:48:08.206287  5203 caffe.cpp:321] ip = -1.1105
I0617 13:48:08.206292  5203 caffe.cpp:321] ip = 0.394778
I0617 13:48:08.206297  5203 caffe.cpp:321] ip = -0.198426
I0617 13:48:08.206301  5203 caffe.cpp:321] ip = -0.884502

发现比没有反向计算，因为不需要求loss，且输出是10个值，与网络结构对应 num_output: 10。

转化自己的raw数据为LMDB数据

对于自己的MNIST数据，使用 /home/junhui/caffe-master/build/examples/mnist/convert_mnist_data.bin 命令，用法如下：

1
2
# 命令 图片文件 对应标签文件 保存位置
convert_mnist_data.bin input_image_file input_label_file output_db_file

那么caffe是如何将所下载的mnist数据集转化为LMDB的，见examples/mnist/create_mnist.sh，分析一下：

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
# 只要这个bash脚本中任何一条命令的返回值不是0，
# 就退出整个脚本，因为默认是有错继续执行。
set -e

# 4个变量，
EXAMPLE=examples/mnist
DATA=data/mnist
BUILD=build/examples/mnist

BACKEND="lmdb"

# 打印内容：
echo "Creating ${BACKEND}..."

# 将之前的DB文件删除
rm -rf $EXAMPLE/mnist_train_${BACKEND}
rm -rf $EXAMPLE/mnist_test_${BACKEND}

# 调用 convert_mnist_data.bin 命令，
$BUILD/convert_mnist_data.bin $DATA/train-images-idx3-ubyte \
  $DATA/train-labels-idx1-ubyte $EXAMPLE/mnist_train_${BACKEND} --backend=${BACKEND}
$BUILD/convert_mnist_data.bin $DATA/t10k-images-idx3-ubyte \
  $DATA/t10k-labels-idx1-ubyte $EXAMPLE/mnist_test_${BACKEND} --backend=${BACKEND}
# 最终在制定的路径中生成 mnist_train_ 和 mnist_test_ 。

echo "Done."

convert_mnist_data.bin 命令对应的源文件在 examples/mnist/convert_mnist_data.cpp，阅读发现其解析过程是根据 train-images-idx3-ubyte 文件的格式的。

下一步：

  1. 如何从LMDB中读取每一条数据？
  2. 将不含标签的数据保存为LMDB？
  3. 写自己的 convert_XXX_data.bin 工具，和bash脚本。