Peta数据集识别性别

数据集预处理

（1）对每一张图片进行resize， resize到特定的大小50*150

（2）rebalance处理，对少数类样本进行随机选择n张进行数据增强之后重新加入到dataset中。

（3）划分训练集和测试集，0.2

（4）对训练集进行数据增强。扩大训练数据量。（操作包括：翻转，滤波等）

#-*- encoding: utf-8 -*-import os, sys, cv2
import numpy as np
import randomimage_cnt = 0
MIN_HEIGHT = 120           #处理的最小尺寸
MIN_WIDTH = 40
targetLabel = []positive_cnt = 0
negative_cnt = 0def readImage( filePath , targetDir ): #制定标签global image_cnt, positive_cnt, negative_cntglobal targetLabelif not os.path.isdir( filePath ):print('{} is not a dir'.format(filePath))return NonelistFile = os.listdir( filePath )labelDict = {}with open( filePath + 'Label.txt', 'r') as reader:for line in reader:lines = line.split()for i in range(1, len(lines)):if lines[i] == 'personalMale': #男设为1label = 1elif lines[i] == 'personalFemale': #女设为0label = 0else:continuelabelDict[lines[0]] = labelbreakfor i in range(len(listFile)):if len(listFile[i]) > 4 and (listFile[i][-4:] == '.bmp' or listFile[i][-4:] == '.jpg' or \listFile[i][-4:] == '.png' or listFile[i][-5:] == '.jpeg'):imageName = filePath + listFile[i]img = cv2.imread( imageName )if not img.data:continueheight, width = img.shape[:2]if height < MIN_HEIGHT or width < MIN_WIDTH:continuefileName = str( image_cnt ) + '.jpeg'identity = listFile[i].find('_') if  identity == -1:identity = len(listFile[i]) idd = listFile[i][:identity]if labelDict.has_key( idd ) :targetLabel.append([ fileName, labelDict[idd]])if labelDict[idd] == 0:negative_cnt += 1else:positive_cnt += 1img = cv2.resize(img, (50, 150), interpolation=cv2.INTER_CUBIC) #60*60cv2.imwrite(targetDir + fileName, img)image_cnt += 1else:print('file {} do not have label'.format(listFile[i]) )####### pyramid operator
def MinAndEnlarge(img, Minus_pixel = 3):   #定义放缩图片大小50*150img = img[(3*Minus_pixel):(150 - 3*Minus_pixel), Minus_pixel:(50 - Minus_pixel), :]img = cv2.resize(img, (50, 150), interpolation = cv2.INTER_CUBIC )return img####### rotate operator
def Flip(img, operator = 1):if operator == 1:img = cv2.flip(img, 1)else:img = cv2.flip(img, 0)return img####### median blurring the image
def Blur(img, kernel_size=5):img = cv2.medianBlur(img, kernel_size)return imgdef saveLabel( targetDir ):      #存储标签global targetLabelwith open(targetDir + 'label.txt', 'w') as writer:for i in range(len(targetLabel)):writer.write( str( targetLabel[i][0] ) + ' ' + str(targetLabel[i][1]) + '\n' )##### ReBalance operator
#######  num (the number of minority class should added)
#######  n_or_p (the label of minority class)
#######  op_chose( 1--symmetrical flip; 0--rotate image; )
def ReBalance( targetDir, num, n_or_p, op_chose = 0):  #平衡正负样本global targetLabel, image_cnt total_sample = len(targetLabel)Contain = {}while 1:if num <= 0:breakkey_id = random.randint(0, total_sample-1)if Contain.has_key( key_id ) or targetLabel[key_id][1] != n_or_p:continueimg = cv2.imread( targetDir + targetLabel[key_id][0] )if op_chose == 0:img = cv2.flip(img, 1) elif op_chose == 1:img = cv2.flip(img, 0)fileName = str(image_cnt) + '.jpeg' cv2.imwrite(targetDir + fileName, img)image_cnt += 1 targetLabel.append([fileName, n_or_p])num -= 1print('Finish add {} images'.format(image_cnt - total_sample)) print('Now the class is balanced and total num is {}'.format(image_cnt)) print('image_cnt is {} and len(_targetLabel_) is {} '.format(image_cnt, len(targetLabel)))def DivideSet( targetDir, trainDir, testDir, test_ratio = 0.20):global targetLabeltotal_sample = len(targetLabel)assert( test_ratio < 1 )test_num = int(test_ratio * total_sample)test_half_num = test_num //2 ; ml_cnt = test_half_num; fm_cnt = test_half_num testLabel = [] ; trainLabel = [] ; testDict = {}while ml_cnt > 0 or fm_cnt > 0:idd = random.randint(0, total_sample-1)if testDict.has_key( targetLabel[idd][0] ):continueif targetLabel[idd][1] == 1 and ml_cnt > 0:img = cv2.imread( targetDir + targetLabel[idd][0] )cv2.imwrite( testDir + targetLabel[idd][0], img )testLabel.append( targetLabel[idd] )testDict[targetLabel[idd][0]] = iddml_cnt -= 1if targetLabel[idd][1] == 0 and fm_cnt > 0:img = cv2.imread( targetDir + targetLabel[idd][0] )cv2.imwrite( testDir + targetLabel[idd][0], img )testLabel.append( targetLabel[idd] )testDict[targetLabel[idd][0]] = iddfm_cnt -= 1for i in range(total_sample):if not testDict.has_key( targetLabel[i][0] ):trainLabel.append( targetLabel[i] )img = cv2.imread( targetDir + targetLabel[i][0] )cv2.imwrite( trainDir + targetLabel[i][0], img )## save the trainset and testsetwith open( trainDir + 'label.txt', 'w') as writer:for i in range(len(trainLabel)):writer.write( str( trainLabel[i][0] ) + ' ' + str(trainLabel[i][1]) + '\n' )   with open( testDir + 'label.txt', 'w') as writer:for i in range(len(testLabel)):writer.write( str(testLabel[i][0]) + ' ' + str(testLabel[i][1]) + '\n')print('has divide into train with {} samples and test with {} samples'.format(len(trainLabel), len(testLabel)) ) return trainLabel, testLabeldef EnlargeTrain( fileDir, targetDir, trainLabel , start_cnt):      #扩大训练集total_sample = len(trainLabel)new_cnt = start_cnt    for i in range(total_sample):img = cv2.imread( fileDir + trainLabel[i][0] )fileLabel = trainLabel[i][1]if not img.data:print('no exits image file {}'.format( fileDir + trainLabel[i][0]) )continue#img1 = MinAndEnlarge(img, 3)fileName = str(new_cnt) + '.jpeg'cv2.imwrite( targetDir + fileName, img1 )new_cnt += 1trainLabel.append( [fileName, fileLabel] )#img2 = Flip(img1)fileName = str(new_cnt) + '.jpeg'cv2.imwrite( targetDir + fileName, img2 )new_cnt += 1trainLabel.append( [fileName, fileLabel] )#img3 = Blur(img, 5)fileName = str(new_cnt) + '.jpeg'cv2.imwrite( targetDir + fileName, img3 )new_cnt += 1trainLabel.append( [fileName, fileLabel] )#img4 = Blur(img1, 5)fileName = str(new_cnt) + '.jpeg'cv2.imwrite( targetDir + fileName, img4 )new_cnt += 1trainLabel.append([fileName, fileLabel])#img5 = Blur(img2, 5)fileName = str(new_cnt) + '.jpeg'cv2.imwrite( targetDir + fileName, img5 )new_cnt += 1trainLabel.append([fileName, fileLabel])print('The total number of training images is {}'.format(new_cnt))with open( targetDir + 'label.txt', 'w') as writer:for i in range(len(trainLabel)):writer.write( str( trainLabel[i][0] ) + ' ' + str(trainLabel[i][1]) + '\n' )print('The trainLabel size is {}'.format(len(trainLabel)) )if __name__ == '__main__':fileHead = 'D:/caffe/caffe-master/data/myself/PETA/PETA dataset/'filePath = ['3DPeS', 'CAVIAR4REID', 'CUHK', 'GRID','MIT', 'PRID']     #原数据集路径，因为太多很慢，只用了一半数据##,'SARC3D','TownCentre', 'VIPeR','i-LID'##] savePath = 'D:/caffe/caffe-master/data/chuliDATA/'                          #处理后的数据和标签存放地址for i in range(len(filePath)):path = fileHead + filePath[i] + '/archive/'print ('runing dataset {}'.format(filePath[i]) ) readImage( path, savePath )print ('The cnt is {}'.format( image_cnt ))#EnlargeData( savePath, savePath )saveLabel( savePath )print( 'we have {} positive labels and {} negative labels '.format( positive_cnt, negative_cnt ))if positive_cnt > negative_cnt:add_num = positive_cnt - negative_cntReBalance( savePath, add_num, 0, 0)else:add_num = negative_cnt - positive_cntReBalance( savePath, add_num, 1, 0)print('The total dataset is in {}'.format(savePath))TrainsavePath = 'D:/caffe/caffe-master/data/myself/petaTrain/'       #训练集存放地址TestsavePath = 'D:/caffe/caffe-master/data/myself/petaTest/'     #测试集存放地址trainLabel, testLabel =  DivideSet(savePath, TrainsavePath, TestsavePath, 0.2 )        #0.2比例划分start_cnt = len(targetLabel)EnlargeTrain( TrainsavePath, TrainsavePath, trainLabel, start_cnt )print('the end')

完成后得到训练集56460，测试集2361
这里图片越大，数据集越大，训练的越慢，还可能爆显存，不得不降低batchsize，图片大小对准确率的影响还不清楚

转换格式

#!/usr/bin/env sh
# Create the imagenet lmdb inputs
# N.B. set the path to the imagenet train + val data dirsEXAMPLE=D:/caffe/caffe-master/data/myself/convertdata
DATA=D:/caffe/caffe-master/data/myself/
TOOLS=D:/caffe/caffe-master/Build/x64/ReleaseTRAIN_DATA_ROOT=D:/caffe/caffe-master/data/myself/petaTrain/
VAL_DATA_ROOT=D:/caffe/caffe-master/data/myself/petaTest/echo "Creating train lmdb..."GLOG_logtostderr=1 $TOOLS/convert_imageset \--shuffle=1 \$TRAIN_DATA_ROOT \$DATA/petaTrain/label.txt \$EXAMPLE/peta_train_lmdbecho "Creating val lmdb..."GLOG_logtostderr=1 $TOOLS/convert_imageset \--shuffle=1 \$VAL_DATA_ROOT \$DATA/petaTest/label.txt \$EXAMPLE/peta_val_lmdbecho "Done."
#需要提前建立文件夹及空标签文件

得到2GB训练集，64M测试集

训练

采用他人的模型，修改描述文件复制到操作文件夹，使用命令行训练，修改模型描述文件如下

name: "petaNet"
layer {name: "peta"type: "Data"top: "data"top: "label"include{phase: TRAIN}data_param{source:"D:/caffe/caffe-master/data/myself/convertdata/peta_train_lmdb"batch_size:32backend: LMDB}
}layer {name: "peta"type: "Data"top: "data"top: "label"include{phase: TEST}data_param{source:"D:/caffe/caffe-master/data/myself/convertdata/peta_val_lmdb"batch_size:100backend: LMDB}
}### ------------layer {name: "conv1"type: "Convolution"bottom: "data"top: "conv1"param {lr_mult: 1}param {lr_mult: 2}convolution_param {num_output: 20kernel_size: 3stride: 1weight_filler {type: "xavier"}bias_filler {type: "constant"}}
}
layer {name: "relu1"type: "ReLU"bottom: "conv1"top: "conv1"
}##-------
# Inception 3a
##-------layer {name: "inc1_conv1"bottom: "conv1"top: "inc1_conv1"type: "Convolution"param { lr_mult: 1 }param { lr_mult: 2 }convolution_param {num_output: 20kernel_size: 7stride: 1weight_filler { type: "xavier" }bias_filler { type: "constant" }}
}
layer {name: "inc1_conv1_relu"type: "ReLU"bottom: "inc1_conv1"top: "inc1_conv1"
}layer {name: "inc1_conv2_1"type: "Convolution"bottom: "conv1"top: "inc1_conv2_1"param { lr_mult: 1 }param { lr_mult: 2 }convolution_param {num_output: 50kernel_size: 3stride: 1weight_filler { type: "xavier" }bias_filler { type: "constant" }}
}
layer {name: "inc1_conv2_1_relu"type: "ReLU"bottom: "inc1_conv2_1"top:    "inc1_conv2_1"
}layer {name: "inc1_conv2_2"type: "Convolution"bottom: "inc1_conv2_1"top:    "inc1_conv2_2"param { lr_mult: 1 }param { lr_mult: 2 }convolution_param {num_output: 50kernel_size: 3stride: 1weight_filler { type: "xavier" }bias_filler { type: "constant" }}
}
layer {name: "inc1_conv2_2_relu"type: "ReLU"bottom: "inc1_conv2_2"top: "inc1_conv2_2"
}layer {name: "inc1_conv2_3"type: "Convolution"bottom: "inc1_conv2_2"top:    "inc1_conv2_3"param { lr_mult: 1 }param { lr_mult: 2 }convolution_param {num_output: 50kernel_size: 3stride: 1weight_filler { type: "xavier" }bias_filler { type: "constant" }}
}
layer {name: "inc1_conv2_3_relu"type: "ReLU"bottom: "inc1_conv2_3"top:    "inc1_conv2_3"
}layer {name: "inc1_conv3_1"type: "Convolution"bottom: "conv1"top:    "inc1_conv3_1"param { lr_mult: 1 }param { lr_mult: 2 }convolution_param {num_output: 20kernel_size: 5stride: 1weight_filler { type: "xavier" }bias_filler { type: "constant" }}
}
layer {name: "inc1_conv3_1_relu"type: "ReLU"bottom: "inc1_conv3_1"top: "inc1_conv3_1"
}layer {name: "inc1_conv3_2"type: "Convolution"bottom: "inc1_conv3_1"top:    "inc1_conv3_2"param { lr_mult: 1 }param { lr_mult: 2 }convolution_param {num_output: 20kernel_size: 3stride: 1weight_filler { type: "xavier" }bias_filler { type: "constant" }}
}
layer {name: "inc1_conv3_2_relu"type: "ReLU"bottom: "inc1_conv3_2"top:    "inc1_conv3_2"
}layer {name: "inc1_concat"type: "Concat"bottom: "inc1_conv1"bottom: "inc1_conv2_3"bottom: "inc1_conv3_2"top:    "inc1_concat"
}#-----end of Inception 3alayer {name: "pool1"type: "Pooling"bottom: "inc1_concat"top: "pool1"pooling_param {pool: MAXkernel_size: 2stride: 2}
}##------
# Inception 2B
##------layer {name: "inc2_conv1_1"type: "Convolution"bottom: "pool1"top: "inc2_conv1_1"param { lr_mult: 1 }param { lr_mult: 2 }convolution_param {num_output: 120kernel_size: 3stride: 1weight_filler { type: "xavier" }bias_filler { type: "constant" }}
}
layer {name: "inc2_conv1_1_relu"type: "ReLU"bottom: "inc2_conv1_1"top: "inc2_conv1_1"
}layer {name: "inc2_conv1_2"type: "Convolution"bottom: "inc2_conv1_1"top:    "inc2_conv1_2"param { lr_mult: 1 }param { lr_mult: 2 }convolution_param {num_output: 120kernel_size: 3stride: 1weight_filler { type: "xavier" }bias_filler { type: "constant" }}
}
layer {name: "inc2_conv1_2_relu"type: "ReLU" bottom: "inc2_conv1_2"top: "inc2_conv1_2"
}layer {name: "inc2_conv2"type: "Convolution"bottom: "pool1"top:    "inc2_conv2"param { lr_mult: 1 }param { lr_mult: 2 }convolution_param {num_output: 120kernel_size: 5stride: 1weight_filler { type: "xavier" }bias_filler { type: "constant" }}
}
layer {name: "inc2_conv2_relu"type: "ReLU"bottom: "inc2_conv2"top:    "inc2_conv2"
}layer {name: "inc2_concat"type: "Concat"bottom: "inc2_conv1_2"bottom: "inc2_conv2"top: "inc2_concat"
}##----end of Inception 2Blayer {name: "pool2"type: "Pooling"bottom: "inc2_concat"top:    "pool2"pooling_param {pool: MAXkernel_size: 2stride: 2}
}layer {name: "fc1"type: "InnerProduct"bottom: "pool2"top:    "fc1"param { lr_mult: 1 }param { lr_mult: 2 }inner_product_param {num_output: 2weight_filler { type: "xavier" }bias_filler { type: "constant" }}
}#### ----------
layer {name: "accuracy"type: "Accuracy"bottom: "fc1"bottom: "label"top:"accuracy"include{phase: TEST}
}layer {name: "loss"type: "SoftmaxWithLoss"bottom: "fc1"bottom: "label"top:"loss"
}

主要修改了全连接层输出个数，输入层来源等。

修改solver如下

# 训练网络描述文件
# The train/test net protocol buffer definition
net: "D:/caffe/caffe-master/data/myself/petaNet-train-test.prototxt"# 测试全部测试集的迭代次数
test_iter: 23
# 每500次迭代测试一次
test_interval: 500
# T基础学习率，上一次梯度更新的权重，权重衰减项，防止过拟合
base_lr: 0.0001
momentum: 0.9
weight_decay: 0.0005
# 学习率变化方式
lr_policy: "inv"
gamma: 0.0001
power: 0.75
# 每100次展示一次准确率
display: 100
# 最大迭代次数
max_iter: 10000
# 每5000次迭代得到一个模型
snapshot: 5000
snapshot_prefix: "D:/caffe/caffe-master/data/myself/Net/"
# 用cpu或gpu
solver_mode: GPU

一开始出现爆显存的问题，方法是把batch-size变小，或者把图片变小，对准确率影响尚不知
之后出现loss不变，始终为87.3365，网上查到原因很多，总结下有以下几种解决办法

降低学习率，提高batchsize，最初为0.01，一直往小调
降低weight 没尝试
输入归一化
查看输入数据、标签是否有异常应该没有数据错误
数据shuffle
batch normalization的设置未尝试

最终改来改去就可以了，大多数情况下应该是学习率问题，以及在网上查看了loss变化与准确率变化情况。

于是最终得到准确率76的模型，若数据集再大一些，应该可以达到80左右，但太慢了，放弃尝试，也可能是学习率比较小导致收敛比较慢，用数据集测试准确率也是76左右。但是准备单张图片测试时，永远预测为0，应该是我的代码写错了，目前没解决。将loss和accuracy生成可视化曲线没解决。