caffe-第一个caffe程序

数据准备1-写入txt

create_train_img_txt.sh
数据来源：github

#/usr/bin/env sh
DATA=/mnt/hgfs/E/data/mouth #数据目录，下面包含0，1两个文件夹,neutral or smile.60x60pixel
DATA_SAVE=. #创建文件保存目录

echo "Creating train.txt...."

rm -rf $DATA_SAVE/train.txt

find $DATA/0/*.jpg | cut -d '/' -f6- | sed "s/$/ 0/" >> $DATA_SAVE/train.txt
find $DATA/1/*.jpg | cut -d '/' -f6- | sed "s/$/ 1/" >> $DATA_SAVE/train.txt

echo "Done!"

• find-将文件路径写入文件中

find /mnt/hgfs/E/data/mouth/0/*.jpg > 0.txt
find /mnt/hgfs/E/data/mouth/1 -name "*.jpg" > 1.txt

• sed-逐行处理文件，"s/$/ 0/" ,s为搜索，在行后面追加空格+0
• cut-截取字节，字符，字块；-d为检查文件夹，/分隔标志， -fn-m分隔的索引

train.txt

---

数据准备2-写入lmdb

create_dataset_lmdb.sh

#!/usr/bin/env sh
# Create the mouth lmdb inputs
# N.B. set the path to the mouthtrain data dirs
set -e #抛出错误提示

CAFFE=/home/wl/software/caffe #caffe安装路径
EXAMPLE=. #lmdb文件保存路径
DATA=. #train.txt路径
TOOLS=$CAFFE/build/tools #一些caffe工具包路径，在caffe安装目录下
TRAIN_DATA_ROOT=/mnt/hgfs/E/data #图片根目录

# Set RESIZE=true to resize the images to 256x256. Leave as false if images have
# already been resized using another tool.
RESIZE=false
if $RESIZE; then
  RESIZE_HEIGHT=256
  RESIZE_WIDTH=256
else
  RESIZE_HEIGHT=0
  RESIZE_WIDTH=0
fi

if [ ! -d "$TRAIN_DATA_ROOT" ]; then
  echo "Error: TRAIN_DATA_ROOT is not a path to a directory: $TRAIN_DATA_ROOT"
  echo "Set the TRAIN_DATA_ROOT variable in create_dataset_lmdb.sh to the path" \
       "where the training data is stored."
  exit 1
fi

echo "Creating train lmdb..."

rm -r mouth_train_lmdb

GLOG_logtostderr=1 $TOOLS/convert_imageset \
    --resize_height=$RESIZE_HEIGHT \
    --resize_width=$RESIZE_WIDTH \
    --shuffle \
    $TRAIN_DATA_ROOT/ \
    $DATA/train.txt \
    $EXAMPLE/mouth_train_lmdb

echo "Creating val lmdb..."

• $TRAIN_DATA_ROOT/后面有个/,不然路径拼接出错

---

编写训练网络

包括数据层，网络结构，loss层
train.prototxt

#train.prototxt
name "mouthNet"
#datalayer
layer{
    name: "data"
    type: "Data"
    top: "data"
    top: "label"
    include{
        phase: TRAIN
    }
    transform_param{
        scale: 0.00390625
    }
    ata_param{
        source: "mouth_train_lmdb"
        batch_size: 1
        backend: LMDB

    }
}
#conv1
layer{
    name: "conv1"
    type: "Convolution"
    bottom: "data"
    top: "conv1"
    param{
        lr_mult: 1
        decay_mult: 0
    }
    param{
        lr_mult: 2
        decay_mult: 0
    }
    convolution_param{
        num_output: 20
        kernel_size: 5
        stride: 1
        weight_filler{
            type: "xavier"
        }
        bias_filler{
            type: "constant"
        }
    }
}
#pool1
layer{
    name: "pool1"
    type: "Pooling"
    bottom: "conv1"
    top: "pool1"
    pooling_param{
        pool: MAX
        kernel_size: 2
        stride: 2
    }
}
#conv2
layer{
    name: "conv2"
    type: "Convolution"
    bottom: "pool1"
    top: "conv2"
    param{
        lr_mult: 1
        decay_mult: 0
    }
    param{
        lr_mult: 2
        decay_mult: 0
    }
    convolution_param{
        num_output: 50
        kernel_size: 5
        stride: 1
        weight_filler{
            type: "xavier"
        }
        bias_filler{
            type: "convolution_param"
        }
    }
}
#pool2
layer{
    name: "pool2"
    type: "Pooling"
    bottom: "conv2"
    top: "pool2"
    pooling_param{
        pool: MAX
        kernel_size: 2
        stride: 2
    }
}
#ip1
layer{
    name: "ip1"
    type: "InnerProduct"
    bottom: "pool2"
    top: "ip1"
    param{
        lr_mult: 1
        decay_mult: 0
    }
    param{
        lr_mult: 2
        decay_mult: 0
    }
    inner_product_param{
        num_output: 500
        weight_filler{
            type: "xavier"
        }
        bias_filler{
            type: "constant"
        }
    }
}
#relu1
layer{
    name: "relu1"
    type: "ReLU"
    bottom: "ip1"
    top: "relu1"
}
#ip2
layer{
    name: "ip2"
    type: "InnerProduct"
    bottom: "relu1"
    top: "ip2"
    param{
        lr_mult: 1
        decay_mult: 0
    }
    param{
        lr_mult: 2
        decay_mult: 0
    }
    inner_product_param{
        num_output:2
        weight_filler{
            type: "xavier"
        }
        bias_filler{
            type: "constant"
        }
    }
}
#loss
layer{
    name: "loss"
    type: "SigmiodCrossEntroyLoss"
    bottom: "ip2"
    bottom: "label"
    top: "loss"
}

• type: "Data", 数据层

1. include TRAIN or TEST,训练阶段加载还是测试阶段加载
2. transform_para 数据变换，scale表示数据缩放，0.00390625=1/255,表示归一化到0-1
3. data_para source数据集文件，batch_size batch size， bcekend 指定数据db形式

---

• type: "Conveolution"-卷积层

1. param lr_mult:1 decay_mult:0指定权重和偏置的学习率和权值衰减
2. convolution_param 指定num_output kernel_size stride weight_filler bias_filler

---

• type: "Pooling" 池化层

1. pooling_param 指定pool池化类型， kernel_size stride, 没有要学习的参数

---

• type: "InnerProduct"-全连接层

1. w,b param
2. num_output wight_filler bias_filler

---

• type: "ReLU"-激活函数

1. 没有参数学习和特殊操作

---

• type: SigmiodCrossEntroyLoss-二分类的loss层

1. sigmoid函数后接交叉熵损失

---

编写训练

-solver.prototxt

# The train net protocol buffer definition
# this follows "ADAM: A METHOD FOR STOCHASTIC OPTIMIZATION"
net: "train.prototxt"
# All parameters are from the cited paper above
base_lr: 0.001
momentum: 0.9
momentum2: 0.999
# since Adam dynamically changes the learning rate, we set the base learning
# rate to a fixed value
lr_policy: "fixed"
# Display every 100 iterations
display: 100
# The maximum number of iterations
max_iter: 100
# snapshot intermediate results
snapshot: 50
snapshot_prefix: "mouth_"
# solver mode: CPU or GPU
type: "Adam"
solver_mode: CPU

---

训练

• train.sh

#/usr/bin/env sh

echo "strating training......"

CAFFE=/home/wl/software
SOLVER=/home/wl/coding/hw

$CAFFE/caffe/build/tools/caffe train -solver=$SOLVER/solver.prototxt

echo "ending training......."

• CAFFE=/home/wl/software SOLVER=/home/wl/coding/hw,赋值时=不能有空格，否则

train.sh: 5: train.sh: CAFFE: not found train.sh: 6: train.sh: SOLVER: not found train.sh: 8: train.sh: /caffe/build/tools/caffe: not found

---

• ERROR1

F1107 22:11:16.345911 13254 data_transformer.cpp:465] Check failed: datum_channels > 0 (0 vs. 0) 
*** Check failure stack trace: ***
    @     0x7f4324e585cd  google::LogMessage::Fail()
    @     0x7f4324e5a433  google::LogMessage::SendToLog()
    @     0x7f4324e5815b  google::LogMessage::Flush()
    @     0x7f4324e5ae1e  google::LogMessageFatal::~LogMessageFatal()
    @     0x7f432550bcb7  caffe::DataTransformer<>::InferBlobShape()
    @     0x7f43254e8f1e  caffe::DataLayer<>::DataLayerSetUp()
    @     0x7f432547bdd9  caffe::BaseDataLayer<>::LayerSetUp()
    @     0x7f432547c22f  caffe::BasePrefetchingDataLayer<>::LayerSetUp()
    @     0x7f43253cd8ec  caffe::Layer<>::SetUp()
    @     0x7f43253b8fda  caffe::Net<>::Init()
    @     0x7f43253b7a30  caffe::Net<>::Net()
    @     0x7f432551473d  caffe::Solver<>::InitTrainNet()
    @     0x7f4325513f6d  caffe::Solver<>::Init()
    @     0x7f4325513a8d  caffe::Solver<>::Solver()
    @     0x7f4325426d64  caffe::SGDSolver<>::SGDSolver()
    @     0x7f432542d51e  caffe::AdamSolver<>::AdamSolver()
    @     0x7f432542e3c1  caffe::Creator_AdamSolver<>()
    @           0x423ec1  caffe::SolverRegistry<>::CreateSolver()
    @           0x41f6c9  train()
    @           0x421b1f  main
    @     0x7f4323dc8830  __libc_start_main
    @           0x41e649  _start
    @              (nil)  (unknown)
Aborted (core dumped)

搜索这里

OK!

• ERROR2

F1107 22:43:10.218010 13542 sigmoid_cross_entropy_loss_layer.cpp:41] Check failed: bottom[0]->count() == bottom[1]->count() (2 vs. 1) SIGMOID_CROSS_ENTROPY_LOSS layer inputs must have the same count.

我的网络最后一层输出确实是num_output: 2, 我以为这样可以…, 改为1，OK！
但时loss层type: "SoftmaxWithLoss",即可以运行

至此第一个caffe程序完成！

---

测试/验证

• train.prototxt -> train_val.prototxt
  前面加上

#val data
layer{
    name: "data"
    type: "Data"
    top: "data"
    top: "label"
    include{
        phase: TEST
    }
    transform_param{
        scale: 0.00390625
    }
    data_param{
        sorce: "mouth_test_lmdb"
        batch_size: 1
        backend: LMDB
    }
}

后面加上

#acc
layer{
    name: "accuracy"
    type: "Accuracy"
    bottom: "ip2"
    bottom: "label"
    top: "accuracy"
    include{
        phase: TEST
    }
}

• solver.prototxt->train_val_solver.prototxt
  加上

#test
# covering the full testing images.
test_iter: 100
# Carry out testing every 50 training iterations.
test_interval: 50

执行train_val.sh训练&验证

• ERROR1

Check failed: label_value < num_labels (1 vs. 1) 

都说和标签是否从0开始有关，我猜应该和我最后输出num_output: 1有关，改为softmax试试…OK!

• ERROR2

Check failed: datum_width == width (61 vs. 60) 

数据集像素可能不是那么一致…
修改create_dataset_ladb.sh

RESIZE=true
if $RESIZE; then
  RESIZE_HEIGHT=60
  RESIZE_WIDTH=60
else
  RESIZE_HEIGHT=0
  RESIZE_WIDTH=0
fi

OK!

---

inference

• c++
• python
  由于python的易用性，学习使用python接口作推断和data层
  参考编译caffe， 重新编译python接口
• 安装python依赖库

sudo apt install python-pip
pip install numpy

总是出错。。。
虚拟机太卡了，换win10写代码吧。