使用caffe训练自己的CNN
现在有这样的一个场景：给你一张行人的矩形图片， 要你识别出该行人的性别特侦。
分析：
（1），行人的姿态各异，变化多端。很难提取图像的特定特征
（2），正常人判别行人的根据是身材比例。（如果是冬天的情况下，行人穿着厚实，性别识别更加难）
solution：
针对难以提取特定特征的图像，可以采用卷积神经网络CNN去自动提取并训练。
数据准备：
采用 PETA数据集，Pedestrain Attribute Recognition At Far Distance。 该数据集一共包含了19000张标记了行人穿着及性别信息的图片。
Peta dataset source url: http://mmlab.ie.cuhk.edu.hk/projects/PETA.html
数据处理：
针对下载解压之后的数据集，采用的流程是：
（1）对每一张图片进行resize， resize到特定的大小（实验中定为50*150），
（2）对正类负类样本的不均衡情况，进行rebalance处理，实验中对少数类样本进行随机选择n张进行data augmentation之后重新加入到dataset中。
（3）划分training set和testing set， 根据train/test ratio将整个数据样本随机分为两部分。
（4）对training set 进行data augmentation 处理。 扩大训练数据量。 （操作包括： 翻转，滤波等）
#!/usr/bin/env python
#-*- encoding: utf-8 -*-
#########
## The python code to preprocess the images and resize them into (50, 150)
## Date: 2016-09-19
#########
import os, sys, cv2
import numpy as np
import random
image_cnt = 0
MIN_HEIGHT = 120
MIN_WIDTH = 40
targetLabel = []
positive_cnt = 0
negative_cnt = 0
def readImage( filePath , targetDir ):
global image_cnt, positive_cnt, negative_cnt
global targetLabel
if not os.path.isdir( filePath ):
print(‘{} is not a dir‘.format(filePath))
return None
listFile = 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‘:
label = 1
elif lines[i] == ‘personalFemale‘:
label = 0
else:
continue
labelDict[lines[0]] = label
break
for 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:
continue
height, width = img.shape[:2]
if height < MIN_HEIGHT or width < MIN_WIDTH:
continue
fileName = 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 += 1
else:
positive_cnt += 1
img = cv2.resize(img, (50, 150), interpolation=cv2.INTER_CUBIC)
cv2.imwrite(targetDir + fileName, img)
image_cnt += 1
else:
print(‘file {} do not have label‘.format(listFile[i]) )
####### pyramid operator
def MinAndEnlarge(img, Minus_pixel = 3):
img = 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 img
def EnlargeData( filePath , targetDir ):
global image_cnt, targetLabel
total_sample = len(targetLabel)
for i in range(total_sample):
img = cv2.imread( filePath + targetLabel[i][0] )
fileLabel = targetLabel[i][1]
if not img.data:
print(‘no exits image file {}‘.format( filePath + targetLabel[i][0]) )
#
img1 = MinAndEnlarge(img, 3)
fileName = str(image_cnt) + ‘.jpeg‘
cv2.imwrite( targetDir + fileName, img1 )
image_cnt += 1
targetLabel.append( [fileName, fileLabel] )
#
img2 = Flip(img1)
fileName = str(image_cnt) + ‘.jpeg‘
cv2.imwrite( targetDir + fileName, img2 )
image_cnt += 1
targetLabel.append( [fileName, fileLabel] )
#
img3 = Blur(img, 5)
fileName = str(image_cnt) + ‘.jpeg‘
cv2.imwrite( targetDir + fileName, img3 )
image_cnt += 1
targetLabel.append( [fileName, fileLabel] )
#
img4 = Blur(img1, 5)
fileName = str(image_cnt) + ‘.jpeg‘
cv2.imwrite( targetDir + fileName, img4 )
image_cnt += 1
targetLabel.append([fileName, fileLabel])
#
img5 = Blur(img2, 5)
fileName = str(image_cnt) + ‘.jpeg‘
cv2.imwrite( targetDir + fileName, img5 )
image_cnt += 1
targetLabel.append([fileName, fileLabel])
print(‘The total number of images is {}‘.format(image_cnt))
def saveLabel( targetDir ):
global targetLabel
with 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):
global targetLabel, image_cnt
total_sample = len(targetLabel)
Contain = {}
while 1:
if num <= 0:
break
key_id = random.randint(0, total_sample-1)
if Contain.has_key( key_id ) or targetLabel[key_id][1] != n_or_p:
continue
img = 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 -= 1
print(‘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 divide( targetDir, trainDir, testDir, test_ratio = 0.20):
global targetLabel
total_sample = len(targetLabel)
assert( test_ratio < 1)
test_num = int(total_sample * test_ratio )
test_half_num = test_num // 2; ml_cnt = 0; fm_cnt = 0
testLabel = [] ; trainLabel = []
for i in range(total_sample):
if ml_cnt < test_half_num and targetLabel[i][1] == 1:
ml_cnt += 1
img = cv2.imread( targetDir + targetLabel[i][0] )
cv2.imwrite( testDir + targetLabel[i][0], img )
testLabel.append(targetLabel[i])
elif fm_cnt < test_half_num and targetLabel[i][1] == 0:
fm_cnt += 1
img = cv2.imread( targetDir + targetLabel[i][0] )
cv2.imwrite( testDir + targetLabel[i][0], img )
testLabel.append(targetLabel[i])
else:
img = cv2.imread( targetDir + targetLabel[i][0] )
cv2.imwrite( trainDir + targetLabel[i][0], img )
trainLabel.append(targetLabel[i])
# train
with 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, testLabel
def DivideSet( targetDir, trainDir, testDir, test_ratio = 0.20):
global targetLabel
total_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] ):
continue
if 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]] = idd
ml_cnt -= 1
if 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]] = idd
fm_cnt -= 1
for 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 testset
with 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, testLabel
def EnlargeTrain( fileDir, targetDir, trainLabel , start_cnt):
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 += 1
trainLabel.append( [fileName, fileLabel] )
#
img2 = Flip(img1)
fileName = str(new_cnt) + ‘.jpeg‘
cv2.imwrite( targetDir + fileName, img2 )
new_cnt += 1
trainLabel.append( [fileName, fileLabel] )
#
img3 = Blur(img, 5)
fileName = str(new_cnt) + ‘.jpeg‘
cv2.imwrite( targetDir + fileName, img3 )
new_cnt += 1
trainLabel.append( [fileName, fileLabel] )
#
img4 = Blur(img1, 5)
fileName = str(new_cnt) + ‘.jpeg‘
cv2.imwrite( targetDir + fileName, img4 )
new_cnt += 1
trainLabel.append([fileName, fileLabel])
#
img5 = Blur(img2, 5)
fileName = str(new_cnt) + ‘.jpeg‘
cv2.imwrite( targetDir + fileName, img5 )
new_cnt += 1
trainLabel.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 = ‘/home/zhangyd/source/PETA_dataset/‘
filePath = [‘3DPeS‘, ‘CAVIAR4REID‘, ‘CUHK‘, ‘GRID‘,‘MIT‘, ‘PRID‘,‘SARC3D‘,‘TownCentre‘, ‘VIPeR‘,‘i-LID‘]
savePath = ‘/home/zhangyd/source/peta/‘
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_cnt
ReBalance( savePath, add_num, 0, 0)
else:
add_num = negative_cnt - positive_cnt
ReBalance( savePath, add_num, 1, 0)
print(‘The total dataset is in {}‘.format(savePath))
TrainsavePath = ‘/home/zhangyd/source/peta_v1/petaTrain/‘
TestsavePath = ‘/home/zhangyd/source/peta_v1/petaTest/‘
trainLabel, testLabel = DivideSet(savePath, TrainsavePath, TestsavePath, 0.2 )
start_cnt = len(targetLabel)
EnlargeTrain( TrainsavePath, TrainsavePath, trainLabel, start_cnt )
print(‘the end‘)
实验
使用caffe的create_lmdb.sh 转换图像数据 成 imbd数据集。
定义 prototxt 等信息。
CNN 的结构是：
训练的参数设置：
# The train/test net protocol buffer definition net: "examples/peta/petanet_train_test.prototxt" # test_iter specifies how many forward passes the test should carry out. # In the case of MNIST, we have test batch size 100 and 100 test iterations, # covering the full 10,000 testing images. test_iter: 100 # Carry out testing every 500 training iterations. test_interval: 500 # The base learning rate, momentum and the weight decay of the network. base_lr: 0.01 momentum: 0.9 weight_decay: 0.0005 # The learning rate policy lr_policy: "inv" gamma: 0.0001 power: 0.75 # Display every 100 iterations display: 100 # The maximum number of iterations max_iter: 10000 # snapshot intermediate results snapshot: 5000 snapshot_prefix: "examples/peta/petanet" # solver mode: CPU or GPU solver_mode: GPU
使用论文《Learned vs. Hand-Crafted Features for Pedestrian Gender Recognition》中的网络结构，取得了较好的训练结果：
I0922 00:07:32.204310 16398 solver.cpp:337] Iteration 10000, Testing net (#0) I0922 00:07:34.001411 16398 solver.cpp:404] Test net output #0: accuracy = 0.8616 I0922 00:07:34.001471 16398 solver.cpp:404] Test net output #1: loss = 0.721973 (* 1 = 0.721973 loss) I0922 00:07:34.001479 16398 solver.cpp:322] Optimization Done. I0922 00:07:34.001485 16398 caffe.cpp:254] Optimization Done.
实验分析：
因为网络不大，网络也比较简单，在GPU下进行训练，消耗的显存大概是几百M，不到1G的显存。网络结构经典，也取得较好的训练结果。
我的拓展： 自己设计的CNN网络
吸取了GoogleNet的网络特征， 引入inception， 重新设计网络。
是两个 inception 组成， 后面加上一个FC层。
其中Snapshot的网络结构prototxt文件是：
name: "petaNet"
layer {
name: "data"
type: "Input"
top: "data"
input_param {
shape: {
dim: 1
dim: 3
dim: 50
dim: 150
}
}
}
### ------------
layer {
name: "conv1"
type: "Convolution"
bottom: "data"
top: "conv1"
param {
lr_mult: 1
}
param {
lr_mult: 2
}
convolution_param {
num_output: 20
kernel_size: 3
stride: 1
weight_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: 20
kernel_size: 7
stride: 1
weight_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: 50
kernel_size: 3
stride: 1
weight_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: 50
kernel_size: 3
stride: 1
weight_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: 50
kernel_size: 3
stride: 1
weight_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: 20
kernel_size: 5
stride: 1
weight_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: 20
kernel_size: 3
stride: 1
weight_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 3a
layer {
name: "pool1"
type: "Pooling"
bottom: "inc1_concat"
top: "pool1"
pooling_param {
pool: MAX
kernel_size: 2
stride: 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: 120
kernel_size: 3
stride: 1
weight_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: 120
kernel_size: 3
stride: 1
weight_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: 120
kernel_size: 5
stride: 1
weight_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 2B
layer {
name: "pool2"
type: "Pooling"
bottom: "inc2_concat"
top: "pool2"
pooling_param {
pool: MAX
kernel_size: 2
stride: 2
}
}
layer {
name: "fc1"
type: "InnerProduct"
bottom: "pool2"
top: "fc1"
param { lr_mult: 1 }
param { lr_mult: 2 }
inner_product_param {
num_output: 2
weight_filler { type: "xavier" }
bias_filler { type: "constant" }
}
}
#### ----------
layer {
name: "prob"
type: "Softmax"
bottom: "fc1"
top: "prob"
}
取得的效果比论文中的网络结构差点， 训练结果是：
I0927 00:11:42.485725 20295 solver.cpp:317] Iteration 10000, loss = 0.0678897 I0927 00:11:42.485771 20295 solver.cpp:337] Iteration 10000, Testing net (#0) I0927 00:12:06.291497 20295 solver.cpp:404] Test net output #0: accuracy = 0.8448 I0927 00:12:06.291554 20295 solver.cpp:404] Test net output #1: loss = 0.614111 (* 1 = 0.614111 loss) I0927 00:12:06.291563 20295 solver.cpp:322] Optimization Done. I0927 00:12:06.291568 20295 caffe.cpp:254] Optimization Done.
实验分析：
因为该网络的组成较为复杂， inception包含着较大的子网络， 因为训练的时候，需要消耗GPU显存为3G多。训练时间也较长些。
reference:
Learned vs. Hand-Crafted Features for Pedestrian Gender Recognition Grigory Antipov,Sid-Ahmed Berrani
文章来自：http://www.cnblogs.com/zhang-yd/p/5966293.html