tag:
python, selenium, PhantomJS, sklearn, BeautifulSoup, caffe
caffe的安装等配置请自行查阅，可以先只编译一个only cpu的
git代码地址：https://github.com/bbfamily/DogJudge
1. 代理获取
爬一些提供免费代理的网站，获取到的代理要根据速度要求等check，
可扩展爬取的网站，这里只简单爬了两个，代理质量一般，也可以用
Tor不过好像也不怎么好使了
from SpiderProxy import SpiderProxyimport ZLog ZLog.init_logging() pxy = SpiderProxy() pxy.spider_proxy360() pxy.spider_xicidaili() pxy.check_proxy() pxy.save_csv() output:211.151.48.60:8080 check ok139.196.108.68:80 check ok110.178.198.55:8888 check ok106.75.128.90:80 check ok60.194.100.51:80 check ok117.57.188.176:81 check ok45.32.19.10:3128 check ok110.181.181.164:8888 check ok39.87.237.90:81 check ok111.206.81.248:80 check ok47.89.53.92:3128 check ok112.87.106.217:81 check ok218.89.69.211:8088 check ok139.59.180.41:8080 check ok124.133.230.254:80 check ok128.199.186.153:8080 check ok192.249.72.148:3128 check ok112.112.70.116:80 check ok128.199.178.73:8080 check ok178.32.153.219:80 check ok79.141.70.78:3128 check ok119.6.136.122:80 check ok46.219.78.221:8081 check ok proxy_list len=23
2. 狗狗分类数据获取
爬虫可设置项：
g_enable_show:是否使用有界面浏览器还是使用PHANTOMJS
g_enable_proxy:浏览器的进程是否启用代理，默认不需要，下载原图一定是使用代理没有开关
g_enable_debug:单进程，单线程调试模式可以debug断点
g_enable_stream使用流下载图片
K_SCROLL_MOVE_DISTANCE = 200 模拟js window下滑距离，增大提高爬取速度
K_SCROLL_SLEEP_TIME = 3
K_COLLECT_PROCESS_CNT = 3 同时启动进程个数
由于使用了线程池控制max线程数，所以就算你提高K_SCROLL_MOVE_DISTANCE，K_SCROLL_SLEEP_TIME也不会有下载速度的提升，
需要修改线程池初始化现在设置了3倍代理数量，具体详看代码：
with ThreadPoolExecutor(max_workers=len(self.back_proxys) * 3) as executor:
默认启动google有界面浏览器了，因为代理质量太差，所以就起了三个进程，如果要启动多个进程在乎效率，代理质量够好，要使用PHANTOMJS
n_jobs = 3if g_enable_debug: n_jobs = 1 parallel = Parallel( n_jobs=n_jobs, verbose=0, pre_dispatch='2*n_jobs') parallel(delayed(do_spider_parallel)(proxy_df, ind, search_name) for ind, search_name in enumerate(search_list))
使用selenium配合BeautifulSoup，requests爬取图片，达到目标数量或者到所有图片停止
具体请参考SpiderBdImg
SpiderBdImg.spider_bd_img([u'拉布拉多', u'哈士奇', u'金毛', u'萨摩耶', u'柯基', u'柴犬', u'边境牧羊犬', u'比格', u'德国牧羊犬', u'杜宾', u'泰迪犬', u'博美', u'巴哥', u'牛头梗'], use_cache=True) output: makedirs ../gen/baidu/image/金毛 makedirs ../gen/baidu/image/哈士奇 makedirs ../gen/baidu/image/拉布拉多 makedirs ../gen/baidu/image/萨摩耶 makedirs ../gen/baidu/image/柯基 makedirs ../gen/baidu/image/柴犬 makedirs ../gen/baidu/image/边境牧羊犬 makedirs ../gen/baidu/image/比格 makedirs ../gen/baidu/image/德国牧羊犬 makedirs ../gen/baidu/image/杜宾 makedirs ../gen/baidu/image/泰迪犬 makedirs ../gen/baidu/image/博美 makedirs ../gen/baidu/image/巴哥 makedirs ../gen/baidu/image/牛头梗
3. 下一步，人工大概扫一下图片，把太过份的删了，不用太仔细，太概扫扫就完事, 这工具其实也是可以自动识别的，先自己扫扫吧
4. 数据标准化
为caffe的lmdb做准备将图片都转换成jpeg，因为作lmdb使用opencv其它格式有问题 包括下载下来的gif，png等等找到图片，辨识真实图片类型，命名真实名称后缀，将非jpeg的转化为jpeg 具体参考ImgStdHelper
运行成功后所有图片为jpeg后缀名称
import ImgStdHelper
ImgStdHelper.std_img_from_root_dir('../gen/baidu/image/', 'jpg')5. 开始训练模型及准备
5.1 生成训练集文件
!../sh/DogType.sh output: mkdir: ../gen/dog_judge: File exists Create train.txt... train.txt Done..
生成如下格式数据，具体参看gen/dog_judge/Train.txt
train_path = '../gen/dog_judge/Train.txt'print open(train_path).read(400) output: 哈士奇/001e5dd0f5aa0959503324336f24a5ea.jpeg 1 哈士奇/001eae03d6f282d1e9f4cb52331d3e20.jpeg 1 哈士奇/0047ea48c765323a53a614d0ed93353b.jpeg 1 哈士奇/006e3bd75b2375149dab9d0323b9fc59.jpeg 1 哈士奇/0084e12ec1c15235a78489a0f4703859.jpeg 1 哈士奇/009724727e40158f5b84a50a7aaaa99b.jpeg 1 哈士奇/00a9d66c72bbed2861f632d07a98db8d.jpeg 1 哈士奇/00dabcba4437f77859b1d8ed37c85360.jpeg 1
生成数字类别对应的label文件
import pandas as pd
class_map = pd.DataFrame(np.array([[1, 2, 3, 4, 5, 6], ['哈士奇', '拉布拉多', '博美', '柴犬', '德国牧羊犬', '杜宾']]).T,
columns=['class', 'name'],
index=np.arange(0, 6))
class_map.to_csv('../gen/class_map.csv', columns=class_map.columns, index=True)5.2 生成val，test集
TrainValSplit 将train的数据集每个类别按照n_folds=10即分成十分，val占一分，train占九份，与scikit等分割参数n_folds用法一样
在gen下重新生成训练数据集，测试数据集，交织测试数据集，这里的test与val数据一样不过，test没有分类标注
def train_val_split(train_path, n_folds=10):
if n_folds <= 1: raise ValueError('n_folds must > 1') with open(train_path, 'r') as f:
lines = f.readlines()
class_dict = defaultdict(list) for line in lines:
cs = line[line.rfind(' '):]
class_dict[cs].append(line)
train = list()
val = list() for cs in class_dict:
cs_len = len(class_dict[cs])
val_cnt = int(cs_len / n_folds)
val.append(class_dict[cs][:val_cnt])
train.append(class_dict[cs][val_cnt:])
val = list(itertools.chain.from_iterable(val))
train = list(itertools.chain.from_iterable(train))
test = [t.split(' ')[0] for t in val]
fn = os.path.dirname(train_path) + '/train_split.txt'
with open(fn, 'w') as f:
f.writelines(train)
fn = os.path.dirname(train_path) + '/val_split.txt'
with open(fn, 'w') as f:
f.writelines(val)
fn = os.path.dirname(train_path) + '/test_split.txt'
with open(fn, 'w') as f:
f.writelines(test)import TrainValSplit
TrainValSplit.train_val_split(train_path, n_folds=10)
train_path = '../gen/dog_judge/train_split.txt'with open(train_path) as f: print 'train set len = {}'.format(len(f.readlines()))
val_path = '../gen/dog_judge/val_split.txt'with open(val_path) as f: print 'val set len = {}'.format(len(f.readlines()))
output:
train set len = 9628
val set len = 10665.2 生成图片lmdb数据库
echo "Begin..." ROOTFOLDER=../gen/baidu/image OUTPUT=../gen/dog_judge rm -rf $OUTPUT/img_train_lmdb /Users/Bailey/caffe/build/tools/convert_imageset --shuffle \ --resize_height=256 --resize_width=256 \ $ROOTFOLDER $OUTPUT/train_split.txt $OUTPUT/img_train_lmdb rm -rf $OUTPUT/img_val_lmdb /Users/Bailey/caffe/build/tools/convert_imageset --shuffle \ --resize_height=256 --resize_width=256 \ $ROOTFOLDER $OUTPUT/val_split.txt $OUTPUT/img_val_lmdb echo "Done.."!../sh/DogLmdb.sh
有些显示Could not open or find file的是如下这张下载就下载残了的，本来就需要干掉
PIL.Image.open('../gen/baidu/image/德国牧羊犬/023ee4e18ebfa4a3db8793e275fae47e.jpeg')
output_25_0.png
5.4 生成去均值mean pb文件
注意需要替换DogMean.sh中caffe的路径文件为你的目录文件MEANBIN=/Users/Bailey/caffe/build/tools/compute_image_mean
!../sh/DogMean.sh oytput: Begin... ../gen/dog_judge/mean.binaryproto ../gen/dog_judge/mean_val.binaryproto Done..
5.5 使用bvlc_googlenet的solver.prototxt，train_val.prototxt训练自己的数据
**
根据训练数据及测试数据的量修改solver.prototxt，train_val.prototxt**
由于测试数据大概1000 －> batch_size=50, test_iter: 20
训练数据大概10000 －> test_interval: 1000
display: 100 snapshot: 5000(其实snapshot大点没事，反正没次crl ＋ c结束时会生成mode), 如过需要多留几个做对比，可调小
可以把test的mirror设置true反正数据不算多
修改DogTrain.sh 中CAFEBIN=/Users/Bailey/caffe/build/tools/caffe为你的caffe路径
修改solver.prototxt，train_val.prototxt中所有绝对路径为你的路径，没法使用相对路径除非想对caffe路径，那样更麻烦
详情请参考solver.prototxt，train_val.prototxt
之后使用!../sh/DogTrain.sh开始训练数据，由于要打太多日志，就不在ipython中运行了，单独启个窗口来, 生成caffemodel
6. 使用生成的模型进行分类
6.1 构造caffe netimport caffe
caffe.set_mode_cpu()
model_def = '../pb/deploy.prototxt'model_weights = '../gen/dog_judge/dog_judge_train_iter_5000.caffemodel'model_mean_file = '../gen/dog_judge/mean.binaryproto'net = caffe.Net(model_def, model_weights, caffe.TEST)
mean_blob = caffe.proto.caffe_pb2.BlobProto()
mean_blob.ParseFromString(open(model_mean_file, 'rb').read())
mean_npy = caffe.io.blobproto_to_array(mean_blob)
mu = mean_npy.mean(2).mean(2)[0]print 'mu = {}'.format(mu)
transformer = caffe.io.Transformer({'data': net.blobs['data'].data.shape})
transformer.set_transpose('data', (2,0,1))
transformer.set_mean('data', mu)
transformer.set_raw_scale('data', 255)
transformer.set_channel_swap('data', (2,1,0))
for layer_name, blob in net.blobs.iteritems(): print layer_name + '\t' + str(blob.data.shape)import numpy as npimport matplotlib.pyplot as pltimport glob
%matplotlib inline
plt.rcParams['figure.figsize'] = (10, 10)主角终于要上场了我家拉布拉多阿布，使用阿布的平时生活照片作为测试看看准确率怎么样
class_map = pd.read_csv('../gen/class_map.csv', index_col=0)class_map
Snip20161002_6.png
predict_dir = '../abu' img_list = glob.glob(predict_dir + '/*.jpeg')
len(img_list)
output:
22
error_prob = []for img in img_list:
image = caffe.io.load_image(img)
transformed_image = transformer.preprocess('data', image)
plt.imshow(image)
plt.show()
net.blobs['data'].data[...] = transformed_image
output = net.forward()
output_prob = output['prob'][0] print 'predicted class is:', class_map[class_map['class'] == output_prob.argmax()].name.values[0] if output_prob.argmax() <> 2:
error_prob.append(img)print 'predicted class is:', class_map[class_map['class'] == output_prob.argmax()].name.values[0]
output_40_2.png
predicted class is: 拉布拉多
output_40_4.png
predicted class is: 拉布拉多
output_40_0.png
predicted class is: 拉布拉多
output_40_6.png
predicted class is: 拉布拉多
output_40_8.png
output_40_10.png
predicted class is: 拉布拉多
output_40_12.png
predicted class is: 德国牧羊犬
output_40_14.png
predicted class is: 博美
output_40_16.png
predicted class is: 拉布拉多
output_40_18.png
predicted class is: 拉布拉多
output_40_20.png
predicted class is: 拉布拉多
output_40_22.png
predicted class is: 拉布拉多
output_40_24.png
predicted class is: 杜宾
output_40_26.png
predicted class is: 拉布拉多
output_40_28.png
predicted class is: 拉布拉多
output_40_30.png
predicted class is: 拉布拉多
output_40_32.png
predicted class is: 拉布拉多
output_40_34.png
predicted class is: 拉布拉多
output_40_36.png
predicted class is: 杜宾
output_40_38.png
predicted class is: 拉布拉多
output_40_40.png
predicted class is: 拉布拉多
output_40_42.png
predicted class is: 拉布拉多
能到80%的查准率其实出乎我预料，在数据不算多，且质量一般的情况下能达到这种效果不得不说caffe确实牛
有些照片比如阿布拉屎那个，躺着睡觉耳朵都立起来那个都判断对了，我还以为得判断成哈士奇呢
accuary = (len(img_list) - len(error_prob))/float(len(img_list)) accuary output: 0.8181818181818182
看一遍分错的这几个，感觉错的rank基本符合正态分布，没什么特别挖掘的
for img in error_prob: try:
image = caffe.io.load_image(img) except Exception: continue
transformed_image = transformer.preprocess('data', image)
plt.imshow(image)
plt.show()
net.blobs['data'].data[...] = transformed_image
output = net.forward()
output_prob = output['prob'][0]
top_inds = output_prob.argsort()[::-1][:6]
for rank, ind in enumerate(top_inds, 1): print 'probabilities rank {} label is {}'.format(rank, class_map[class_map['class']==ind].name.values[0])print 'probabilities rank {} label is {}'.format(rank, class_map[class_map['class']==ind].name.values[0])
output_40_12.png
probabilities rank 1 label is 德国牧羊犬 probabilities rank 2 label is 杜宾 probabilities rank 3 label is 拉布拉多 probabilities rank 4 label is 柴犬 probabilities rank 5 label is 博美 probabilities rank 6 label is 哈士奇
output_40_14.png
probabilities rank 1 label is 博美 probabilities rank 2 label is 柴犬 probabilities rank 3 label is 拉布拉多 probabilities rank 4 label is 哈士奇 probabilities rank 5 label is 杜宾 probabilities rank 6 label is 德国牧羊犬
output_40_24.png
probabilities rank 1 label is 杜宾 probabilities rank 2 label is 德国牧羊犬 probabilities rank 3 label is 柴犬 probabilities rank 4 label is 哈士奇 probabilities rank 5 label is 拉布拉多 probabilities rank 6 label is 博美
output_40_36.png
probabilities rank 1 label is 杜宾 probabilities rank 2 label is 拉布拉多 probabilities rank 3 label is 德国牧羊犬 probabilities rank 4 label is 柴犬 probabilities rank 5 label is 博美 probabilities rank 6 label is 哈士奇
就写到这里吧，还拿阿布玩的照片分了两类一类是在草地玩， 一类是在水里玩，训练了模型后测试发现准确率
也相当高，说明针对小数据集，caffe确实也可以工作的不错
作者：abu量化
链接：https://www.jianshu.com/p/6942241d4ad9
随时随地看视频
