转载于：http://www.cnblogs.com/ansang/p/8583122.html
上一个代码只能实现小数据的读取与训练，在大数据训练的情况下。会造内存紧张，于是我根据keras的官方文档，对上一个代码进行了改进。
用keras实现人脸关键点检测
数据集：https://pan.baidu.com/s/1cnAxJJmN9nQUVYj8w0WocA
第一步：准备好需要的库
- tensorflow 1.4.0
- h5py 2.7.0
- hdf5 1.8.15.1
- Keras 2.0.8
- opencv-python 3.3.0
- numpy 1.13.3+mkl
第二步：准备数据集：
我对每一张图像进行了剪裁，使图像的大小为178*178的正方形。
并且对于原有的lable进行了优化
第三步：将图片和标签转成numpy array格式：
参数
1 trainpath = 'E:/pycode/facial-keypoints-master/data/50000train/'
2 testpath = 'E:/pycode/facial-keypoints-master/data/50000test/'
3 imgsize = 178
4 train_samples =40000
5 test_samples = 200
6 batch_size = 32
1 def __data_label__(path):
2 f = open(path + "lable-40.txt", "r")
3 j = 0
4 i = -1
5 datalist = []
6 labellist = []
7 while True:
8
9 for line in f.readlines():
10 i += 1
11 j += 1
12 a = line.replace("\n", "")
13 b = a.split(",")
14 lable = b[1:]
15 # print(b[1:])
16 #对标签进行归一化（不归一化也行）
17 # for num in b[1:]:
18 # lab = int(num) / 255.0
19 # labellist.append(lab)
20 # lab = labellist[i * 10:j * 10]
21 imgname = path + b[0]
22 images = load_img(imgname)
23 images = img_to_array(images).astype('float32')
24 # 对图片进行归一化（不归一化也行）
25 # images /= 255.0
26 image = np.expand_dims(images, axis=0)
27 lables = np.array(lable)
28
29 # lable =keras.utils.np_utils.to_categorical(lable)
30 # lable = np.expand_dims(lable, axis=0)
31 lable = lables.reshape(1, 10)
32 #这里使用了生成器
33 yield (image,lable)
第四步：搭建网络：
这里使用非常简单的网络
1 def __CNN__(self):
2 model = Sequential()#178*178*3
3 model.add(Conv2D(32, (3, 3), input_shape=(imgsize, imgsize, 3)))
4 model.add(Activation('relu'))
5 model.add(MaxPooling2D(pool_size=(2, 2)))
6
7 model.add(Conv2D(32, (3, 3)))
8 model.add(Activation('relu'))
9 model.add(MaxPooling2D(pool_size=(2, 2)))
10
11 model.add(Conv2D(64, (3, 3)))
12 model.add(Activation('relu'))
13 model.add(MaxPooling2D(pool_size=(2, 2)))
14
15 model.add(Flatten())
16 model.add(Dense(64))
17 model.add(Activation('relu'))
18 model.add(Dropout(0.5))
19 model.add(Dense(10))
20 return model
21 #因为是回归问题，抛弃了softmax
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
conv2d_1 (Conv2D) (None, 176, 176, 32) 896
_________________________________________________________________
activation_1 (Activation) (None, 176, 176, 32) 0
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 88, 88, 32) 0
_________________________________________________________________
conv2d_2 (Conv2D) (None, 86, 86, 32) 9248
_________________________________________________________________
activation_2 (Activation) (None, 86, 86, 32) 0
_________________________________________________________________
max_pooling2d_2 (MaxPooling2 (None, 43, 43, 32) 0
_________________________________________________________________
conv2d_3 (Conv2D) (None, 41, 41, 64) 18496
_________________________________________________________________
activation_3 (Activation) (None, 41, 41, 64) 0
_________________________________________________________________
max_pooling2d_3 (MaxPooling2 (None, 20, 20, 64) 0
_________________________________________________________________
flatten_1 (Flatten) (None, 25600) 0
_________________________________________________________________
dense_1 (Dense) (None, 64) 1638464
_________________________________________________________________
activation_4 (Activation) (None, 64) 0
_________________________________________________________________
dropout_1 (Dropout) (None, 64) 0
_________________________________________________________________
dense_2 (Dense) (None, 10) 650
=================================================================
Total params: 1,667,754
Trainable params: 1,667,754
Non-trainable params: 0
_________________________________________________________________
第五步：训练网络：
1 def train(model):
2 # print(lable.shape)
3 model.compile(loss='mse', optimizer='adam')
4 # optimizer = SGD(lr=0.03, momentum=0.9, nesterov=True)
5 # model.compile(loss='mse', optimizer=optimizer, metrics=['accuracy'])
6 epoch_num = 14
7 learning_rate = np.linspace(0.03, 0.01, epoch_num)
8 change_lr = LearningRateScheduler(lambda epoch: float(learning_rate[epoch]))
9 early_stop = EarlyStopping(monitor='val_loss', patience=20, verbose=1, mode='auto')
10 check_point = ModelCheckpoint('CNN_model_final.h5', monitor='val_loss', verbose=0, save_best_only=True,
11 save_weights_only=False, mode='auto', period=1)
12
13 model.fit_generator(__data_label__(trainpath),callbacks=[check_point,early_stop,change_lr],samples_per_epoch=int(train_samples // batch_size),
14 epochs=epoch_num,validation_steps = int(test_samples // batch_size),validation_data=__data_label__(testpath))
15
16 # model.fit(traindata, trainlabel, batch_size=32, epochs=50,
17 # validation_data=(testdata, testlabel))
18 model.evaluate_generator(__data_label__(testpath),steps=10)
19
20 def save(model, file_path=FILE_PATH):
21 print('Model Saved.')
22 model.save_weights(file_path)
23
24
25 def predict(model,image):
26 # 预测样本分类
27 image = cv2.resize(image, (imgsize, imgsize))
28 image.astype('float32')
29 image /= 255
30
31 #归一化
32 result = model.predict(image)
33 result = result*1000+20
34
35 print(result)
36 return result
使用了fit_generator这一方法，加入了learning_rate,LearningRateScheduler,early_stop等参数。
第六步：图像验证
1 import tes_main
2 from keras.preprocessing.image import load_img, img_to_array
3 import numpy as np
4 import cv2
5 FILE_PATH = 'E:\\pycode\\facial-keypoints-master\\code\\CNN_model_final.h5'
6 imgsize =178
7 def point(img,x, y):
8 cv2.circle(img, (x, y), 1, (0, 0, 255), 10)
9
10 Model = tes_main.Model()
11 model = Model.__CNN__()
12 Model.load(model,FILE_PATH)
13 img = []
14 # path = "D:\\Users\\a\\Pictures\\face_landmark_data\data\\test\\000803.jpg"
15 path = "E:\pycode\\facial-keypoints-master\data\\50000test\\049971.jpg"
16 # image = load_img(path)
17 # img.append(img_to_array(image))
18 # img_data = np.array(img)
19 imgs = cv2.imread(path)
20 # img_datas = np.reshape(imgs,(imgsize, imgsize,3))
21 image = cv2.resize(imgs, (imgsize, imgsize))
22 rects = Model.predict(model,imgs)
23
24 for x, y, w, h, a,b,c,d,e,f in rects:
25 point(image,x,y)
26 point(image,w, h)
27 point(image,a,b)
28 point(image,c,d)
29 point(image,e,f)
30
31 cv2.imshow('img', image)
32 cv2.waitKey(0)
33 cv2.destroyAllWindows()
完整代码如下
1 from tensorflow.contrib.keras.api.keras.preprocessing.image import ImageDataGenerator,img_to_array
2 from keras.models import Sequential
3 from keras.layers.core import Dense, Dropout, Activation, Flatten
4 from keras.layers.advanced_activations import PReLU
5 from keras.layers.convolutional import Conv2D, MaxPooling2D,ZeroPadding2D
6 from keras.preprocessing.image import load_img, img_to_array
7 from keras.optimizers import SGD
8 import numpy as np
9 import cv2
10 from keras.callbacks import *
11 import keras
12
13 FILE_PATH = 'E:\\pycode\\facial-keypoints-master\\code\\CNN_model_final.h5'
14 trainpath = 'E:/pycode/facial-keypoints-master/data/50000train/'
15 testpath = 'E:/pycode/facial-keypoints-master/data/50000test/'
16 imgsize = 178
17 train_samples =40000
18 test_samples = 200
19 batch_size = 32
20 def __data_label__(path):
21 f = open(path + "lable-40.txt", "r")
22 j = 0
23 i = -1
24 datalist = []
25 labellist = []
26 while True:
27
28 for line in f.readlines():
29 i += 1
30 j += 1
31 a = line.replace("\n", "")
32 b = a.split(",")
33 lable = b[1:]
34 # print(b[1:])
35 #对标签进行归一化（不归一化也行）
36 # for num in b[1:]:
37 # lab = int(num) / 255.0
38 # labellist.append(lab)
39 # lab = labellist[i * 10:j * 10]
40 imgname = path + b[0]
41 images = load_img(imgname)
42 images = img_to_array(images).astype('float32')
43 # 对图片进行归一化（不归一化也行）
44 # images /= 255.0
45 image = np.expand_dims(images, axis=0)
46 lables = np.array(lable)
47
48 # lable =keras.utils.np_utils.to_categorical(lable)
49 # lable = np.expand_dims(lable, axis=0)
50 lable = lables.reshape(1, 10)
51
52 yield (image,lable)
53
54 ###############:
55
56 # 开始建立CNN模型
57 ###############
58
59 # 生成一个model
60 class Model(object):
61 def __CNN__(self):
62 model = Sequential()#218*178*3
63 model.add(Conv2D(32, (3, 3), input_shape=(imgsize, imgsize, 3)))
64 model.add(Activation('relu'))
65 model.add(MaxPooling2D(pool_size=(2, 2)))
66
67 model.add(Conv2D(32, (3, 3)))
68 model.add(Activation('relu'))
69 model.add(MaxPooling2D(pool_size=(2, 2)))
70
71 model.add(Conv2D(64, (3, 3)))
72 model.add(Activation('relu'))
73 model.add(MaxPooling2D(pool_size=(2, 2)))
74
75 model.add(Flatten())
76 model.add(Dense(64))
77 model.add(Activation('relu'))
78 model.add(Dropout(0.5))
79 model.add(Dense(10))
80 model.summary()
81 return model
82
83
84 def train(self,model):
85 # print(lable.shape)
86 model.compile(loss='mse', optimizer='adam', metrics=['accuracy'])
87 # optimizer = SGD(lr=0.03, momentum=0.9, nesterov=True)
88 # model.compile(loss='mse', optimizer=optimizer, metrics=['accuracy'])
89 epoch_num = 10
90 learning_rate = np.linspace(0.03, 0.01, epoch_num)
91 change_lr = LearningRateScheduler(lambda epoch: float(learning_rate[epoch]))
92 early_stop = EarlyStopping(monitor='val_loss', patience=20, verbose=1, mode='auto')
93 check_point = ModelCheckpoint('CNN_model_final.h5', monitor='val_loss', verbose=0, save_best_only=True,
94 save_weights_only=False, mode='auto', period=1)
95
96 model.fit_generator(__data_label__(trainpath),callbacks=[check_point,early_stop,change_lr],samples_per_epoch=int(train_samples // batch_size),
97 epochs=epoch_num,validation_steps = int(test_samples // batch_size),validation_data=__data_label__(testpath))
98
99 # model.fit(traindata, trainlabel, batch_size=32, epochs=50,
100 # validation_data=(testdata, testlabel))
101 model.evaluate_generator(__data_label__(testpath))
102
103 def save(self,model, file_path=FILE_PATH):
104 print('Model Saved.')
105 model.save_weights(file_path)
106
107 def load(self,model, file_path=FILE_PATH):
108 print('Model Loaded.')
109 model.load_weights(file_path)
110
111 def predict(self,model,image):
112 # 预测样本分类
113 print(image.shape)
114 image = cv2.resize(image, (imgsize, imgsize))
115 image.astype('float32')
116 image = np.expand_dims(image, axis=0)
117
118 #归一化
119 result = model.predict(image)
120
121 print(result)
122 return result