def load(self,
img_rows=IMAGE_SIZE,
img_cols=IMAGE_SIZE,
img_channels=3,
nb_classes=2):
# 加载数据集到内存
#images, labels = load_dataset(self.path_name)
#train_images, valid_images, train_labels, valid_labels = \
# train_test_split(images, labels, test_size=0.3, \
# random_state=random.randint(0, 100))
#print(train_images.shape)
#print(valid_images.shape)
train_images, train_labels = load_dataset(
"genderdetectionfacedata/train")
test_images, test_labels = load_dataset("genderdetectionfacedata/test")
valid_images, valid_labels = load_dataset(
"genderdetectionfacedata/valid")
#print(test_images.shape)
self.input_shape = (train_images[0].shape)
# 输出训练集、验证集、测试集的数量
print(train_images.shape[0], 'train samples')
print(valid_images.shape[0], 'valid samples')
print(test_images.shape[0], 'test samples')
# 我们的模型使用categorical_crossentropy作为损失函数，因此需要根据类别数量nb_classes将
# 类别标签进行one-hot编码使其向量化，在这里我们的类别只有两种，经过转化后标签数据变为二维
train_labels = np_utils.to_categorical(train_labels, nb_classes)
#print(train_labels[0])
valid_labels = np_utils.to_categorical(valid_labels, nb_classes)
test_labels = np_utils.to_categorical(test_labels, nb_classes)
# 像素数据浮点化以便归一化
train_images = train_images.astype('float32')
valid_images = valid_images.astype('float32')
test_images = test_images.astype('float32')
# 将其归一化,图像的各像素值归一化到0~1区间
train_images /= 255
valid_images /= 255
test_images /= 255
self.train_images = train_images
self.valid_images = valid_images
self.test_images = test_images
self.train_labels = train_labels
self.valid_labels = valid_labels
self.test_labels = test_labels
def load(self, img_rows=size, img_cols=size, img_channels=3, nb_classes=2):
# 加载数据集到内存
images, labels = load_dataset(self.path_name)
train_images, valid_images, train_labels, valid_labels = train_test_split(
images, labels, test_size=0.2, random_state=random.randint(0, 100))
_, test_images, _, test_labels = train_test_split(
images, labels, test_size=0.2, random_state=random.randint(0, 100))
# 当前的维度顺序如果是"th",则输入图片数据时的顺序为：channels,rows,cols 否则：rows,cols,channels
# 这部分代码就是根据keras库要求的维度顺序重组训练数据集
if K.image_dim_ordering() == 'th':
train_images = train_images.reshape(train_images.shape[0],
img_channels, img_rows,
img_cols)
valid_images = valid_images.reshape(valid_images.shape[0],
img_channels, img_rows,
img_cols)
test_images = test_images.reshape(test_images.shape[0],
img_channels, img_rows, img_cols)
self.input_shape = (img_channels, img_rows, img_cols)
else:
train_images = train_images.reshape(train_images.shape[0],
img_rows, img_cols,
img_channels)
valid_images = valid_images.reshape(valid_images.shape[0],
img_rows, img_cols,
img_channels)
test_images = test_images.reshape(test_images.shape[0], img_rows,
img_cols, img_channels)
self.input_shape = (img_rows, img_cols, img_channels)
# 输出训练集、验证集、测试集的数量
print(train_images.shape[0], 'train samples')
print(valid_images.shape[0], 'valid samples')
print(test_images.shape[0], 'test samples')
# 我们的模型使用catergorical_crossentropy作为损失函数，因此需要根据类别数量nb_classes
# 将类别标签进行one-hot编码使其向量化，在这里我们的类别只有两种，经过转化后标签数据变为二维
train_labels = np_utils.to_categorical(train_labels, nb_classes)
valid_labels = np_utils.to_categorical(valid_labels, nb_classes)
test_labels = np_utils.to_categorical(test_labels, nb_classes)
# 像素数据浮点化以便归一化
train_images = train_images.astype('float32')
valid_images = valid_images.astype('float32')
test_images = test_images.astype('float32')
# 将其归一化，图像的各像素值归一化到0~1区间
train_images /= 255
valid_images /= 255
test_images /= 255
self.train_images = train_images
self.valid_images = valid_images
self.test_images = test_images
self.train_labels = train_labels
self.valid_labels = valid_labels
self.test_labels = test_labels
def load(self,
img_rows=IMAGE_SIZE,
img_cols=IMAGE_SIZE,
img_channels=3,
nb_classes=2):
images, labels = load_dataset(self.path_name)
train_images, valid_images, train_labels, valid_labels = train_test_split(
images, labels, test_size=0.3, random_state=random.randint(0, 100))
_, test_images, _, test_labels = train_test_split(
images, labels, test_size=0.5, random_state=random.randint(0, 100))
if K.image_dim_ordering() == 'th':
train_images = train_images.reshape(train_images.shape[0],
img_channels, img_rows,
img_cols)
valid_images = valid_images.reshape(valid_images.shape[0],
img_channels, img_rows,
img_cols)
test_images = test_images.reshape(test_images.shape[0],
img_channels, img_rows, img_cols)
self.input_shape = (img_channels, img_rows, img_cols)
else:
train_images = train_images.reshape(train_images.shape[0],
img_rows, img_cols,
img_channels)
valid_images = valid_images.reshape(valid_images.shape[0],
img_rows, img_cols,
img_channels)
test_images = test_images.reshape(test_images.shape[0], img_rows,
img_cols, img_channels)
self.input_shape = (img_rows, img_cols, img_channels)
print(train_images.shape[0], 'train samples')
print(valid_images.shape[0], 'valid samples')
print(test_images.shape[0], 'test samples')
train_labels = np_utils.to_categorical(train_labels, nb_classes)
valid_labels = np_utils.to_categorical(valid_labels, nb_classes)
test_labels = np_utils.to_categorical(test_labels, nb_classes)
train_images = train_images.astype('float32')
valid_images = valid_images.astype('float32')
test_images = test_images.astype('float32')
train_images /= 255
valid_images /= 255
test_images /= 255
self.train_images = train_images
self.valid_images = valid_images
self.test_images = test_images
self.train_labels = train_labels
self.valid_labels = valid_labels
self.test_labels = test_labels
def load(self, img_rows = IMAGE_SIZE, img_cols = IMAGE_SIZE,
img_channels = 3):
# 加载数据集到内存
images, labels , nb_classes= load_dataset(self.path_name)
# 导入了sklearn库的交叉验证模块，利用函数train_test_split()来划分训练集和验证集
# 划分出了30%的数据用于验证，70%用于训练模型
train_images, valid_images, train_labels, valid_labels = train_test_split(images,\
labels, test_size = 0.3, random_state = random.randint(0, 100))
_, test_images, _, test_labels = train_test_split(images, labels, test_size = 0.5,\
random_state = random.randint(0, 100))
# 当前的维度顺序如果为'channels_first'，则输入图片数据时的顺序为：channels,rows,cols，否则:rows,cols,channels
# 这部分代码就是根据keras库要求的维度顺序重组训练数据集
if K.image_data_format() == 'channels_first':
train_images = train_images.reshape(train_images.shape[0], img_channels, img_rows, img_cols)
valid_images = valid_images.reshape(valid_images.shape[0], img_channels, img_rows, img_cols)
test_images = test_images.reshape(test_images.shape[0], img_channels, img_rows, img_cols)
self.input_shape = (img_channels, img_rows, img_cols)
else:
train_images = train_images.reshape(train_images.shape[0], img_rows, img_cols, img_channels)
valid_images = valid_images.reshape(valid_images.shape[0], img_rows, img_cols, img_channels)
test_images = test_images.reshape(test_images.shape[0], img_rows, img_cols, img_channels)
self.input_shape = (img_rows, img_cols, img_channels)
# 输出训练集、验证集、测试集的数量
print(train_images.shape[0], 'train samples')
print(valid_images.shape[0], 'valid samples')
print(test_images.shape[0], 'test samples')
# 我们的模型使用categorical_crossentropy作为损失函数，因此需要根据类别数量nb_classes将
# 类别标签进行one-hot编码使其向量化，在这里我们的类别只有两种，经过转化后标签数据变为二维
train_labels = np_utils.to_categorical(train_labels, nb_classes)
valid_labels = np_utils.to_categorical(valid_labels, nb_classes)
test_labels = np_utils.to_categorical(test_labels, nb_classes)
# 像素数据浮点化以便归一化
train_images = train_images.astype('float32')
valid_images = valid_images.astype('float32')
test_images = test_images.astype('float32')
# 将其归一化,图像的各像素值归一化到0~1区间，数据集先浮点后归一化的目的是提升网络收敛速度，
# 减少训练时间，同时适应值域在（0,1）之间的激活函数，增大区分度
train_images /= 255
valid_images /= 255
test_images /= 255
self.train_images = train_images
self.valid_images = valid_images
self.test_images = test_images
self.train_labels = train_labels
self.valid_labels = valid_labels
self.test_labels = test_labels
def load(self, img_rows = IMAGE_SIZE, img_cols = IMAGE_SIZE, img_channels = 3, model = facenet):
images, labels = load_dataset(self.path_name)
X_embedding = img_to_encoding(images, model)
print('X_train shape', X_embedding.shape)
print('y_train shape', labels.shape)
print(X_embedding.shape[0], 'train samples')
self.X_train = X_embedding
self.y_train = labels
def load(self, img_rows=IMAGE_SIZE, img_cols=IMAGE_SIZE,
img_channels=3, nb_classes=2):
# 加载数据集到内存
images, labels = load_dataset(self.path_name)
train_images, valid_images, train_labels, valid_labels = train_test_split(images, labels, test_size=0.3,
random_state=random.randint(0, 100))
_, test_images, _, test_labels = train_test_split(images, labels, test_size=0.5,
random_state=random.randint(0, 100))
# 通道 行 列顺序
# theano 作为后端：channels,rows,cols TensorFlow作为后端:rows,cols,channels
if K.image_dim_ordering() == 'th':
train_images = train_images.reshape(train_images.shape[0], img_channels, img_rows, img_cols)
valid_images = valid_images.reshape(valid_images.shape[0], img_channels, img_rows, img_cols)
test_images = test_images.reshape(test_images.shape[0], img_channels, img_rows, img_cols)
self.input_shape = (img_channels, img_rows, img_cols)
else:
train_images = train_images.reshape(train_images.shape[0], img_rows, img_cols, img_channels)
valid_images = valid_images.reshape(valid_images.shape[0], img_rows, img_cols, img_channels)
test_images = test_images.reshape(test_images.shape[0], img_rows, img_cols, img_channels)
self.input_shape = (img_rows, img_cols, img_channels)
# 输出训练集、验证集、测试集的数量
print(train_images.shape[0], '训练数据')
print(valid_images.shape[0], '验证数据')
print(test_images.shape[0], '测试数据')
# 将类别标签进行one-hot编码使其向量化，两种类别转为二维数组
train_labels = np_utils.to_categorical(train_labels, nb_classes)
valid_labels = np_utils.to_categorical(valid_labels, nb_classes)
test_labels = np_utils.to_categorical(test_labels, nb_classes)
# 像素数据浮点化以便归一化
train_images = train_images.astype('float32')
valid_images = valid_images.astype('float32')
test_images = test_images.astype('float32')
# 将其归一化,图像的各像素值归一化到0~1区间
train_images /= 255
valid_images /= 255
test_images /= 255
self.train_images = train_images
self.valid_images = valid_images
self.test_images = test_images
self.train_labels = train_labels
self.valid_labels = valid_labels
self.test_labels = test_labels
def load(self, img_rows = IMAGE_SIZE, img_cols = IMAGE_SIZE,
img_channels = 3, nb_classes = 2):
#加在數據集到内存
images, labels = load_dataset(self.path_name)
train_images, valid_images, train_labels, valid_labels = train_test_split(images, labels, test_size = 0.4, random_state = random.randint(0, 100))
_, test_images, _, test_labels = train_test_split(images, labels, test_size = 0.3, random_state = random.randint(0, 100))
#當前的維度順序如果為'th'，則輸入照片的順序為：channels,rows,cols，否則:rows,cols,channels
#這部分代碼就是根據keras酷要求的維度順序重組訓練數集
if K.image_dim_ordering() == 'th':
train_images = train_images.reshape(train_images.shape[0], img_channels, img_rows, img_cols)
valid_images = valid_images.reshape(valid_images.shape[0], img_channels, img_rows, img_cols)
test_images = test_images.reshape(test_images.shape[0], img_channels, img_rows, img_cols)
self.input_shape = (img_channels, img_rows, img_cols)
else:
train_images = train_images.reshape(train_images.shape[0], img_rows, img_cols, img_channels)
valid_images = valid_images.reshape(valid_images.shape[0], img_rows, img_cols, img_channels)
test_images = test_images.reshape(test_images.shape[0], img_rows, img_cols, img_channels)
self.input_shape = (img_rows, img_cols, img_channels)
#輸出訓練集、驗證集、測試集的数量
print(train_images.shape[0], 'train samples')
print(valid_images.shape[0], 'valid samples')
print(test_images.shape[0], 'test samples')
#我们的模型使用categorical_crossentropy作為損失函數，因此需要根據類別數量nb_classes將
#類別標前進行one-hot编码使其向量化，在這裡我们的類別只有兩種，經過轉化後標籤數據變為二維
train_labels = np_utils.to_categorical(train_labels, nb_classes)
valid_labels = np_utils.to_categorical(valid_labels, nb_classes)
test_labels = np_utils.to_categorical(test_labels, nb_classes)
#像素數據浮點化以便歸一化
train_images = train_images.astype('float32')
valid_images = valid_images.astype('float32')
test_images = test_images.astype('float32')
#將其歸一化,圖像的各像素值歸一化到0~1區間
train_images /= 255
valid_images /= 255
test_images /= 255
self.train_images = train_images
self.valid_images = valid_images
self.test_images = test_images
self.train_labels = train_labels
self.valid_labels = valid_labels
self.test_labels = test_labels
def load(self,
img_rows=IMAGE_SIZE,
img_cols=IMAGE_SIZE,
img_channels=3,
model=facenet):
# 加载数据集到内存
images, labels = load_dataset(self.path_name)
# 生成128维特征向量
X_embedding = img_to_encoding(
images, model
) # 考虑这里分批执行，否则可能内存不够，这里在img_to_encoding函数里通过predict的batch_size参数实现
# 输出训练集、验证集和测试集的数量
print('X_train shape', X_embedding.shape)
print('y_train shape', labels.shape)
print(X_embedding.shape[0], 'train samples')
# 这里对X_train就不再进一步normalization了，因为已经在facenet里有了l2_norm
self.X_train = X_embedding
self.y_train = labels
def load(self,
img_rows=IMAGE_SIZE,
img_cols=IMAGE_SIZE,
img_channels=3,
nb_classes=3):
# 加载数据集到内存
images, labels = load_dataset(self.path_name)
#train_test_split()函数会按照训练集特征数据（这里就是图像数据）、测试集特征数据、训练集标签、测试集标签的顺序返回各数据集
#train_test_split是交叉验证中常用的函数，功能是从样本中随机的按比例选区trian_data和testdata
train_images, valid_images, train_labels, valid_labels = train_test_split(
images, labels, test_size=0.3, random_state=random.randint(0, 100))
_, test_images, _, test_labels = train_test_split(
images, labels, test_size=0.5, random_state=random.randint(0, 100))
# 后端系统决定图像数据输入cnn网络时的维度顺序，当前的维度顺序如果为'th'(theano)，则输入图片数据时的顺序为：channels,rows,cols，否则:rows,cols,channels(TF)
# 这部分代码就是根据keras库要求的维度顺序重组训练数据集
if K.image_dim_ordering() == 'th':
train_images = train_images.reshape(train_images.shape[0],
img_channels, img_rows,
img_cols)
valid_images = valid_images.reshape(valid_images.shape[0],
img_channels, img_rows,
img_cols)
test_images = test_images.reshape(test_images.shape[0],
img_channels, img_rows, img_cols)
self.input_shape = (img_channels, img_rows, img_cols)
else:
#reshape函数重新调整向量维度
train_images = train_images.reshape(train_images.shape[0],
img_rows, img_cols,
img_channels)
valid_images = valid_images.reshape(valid_images.shape[0],
img_rows, img_cols,
img_channels)
test_images = test_images.reshape(test_images.shape[0], img_rows,
img_cols, img_channels)
self.input_shape = (img_rows, img_cols, img_channels)
# 输出训练集、验证集、测试集的数量
print(train_images.shape[0], 'train samples')
print(valid_images.shape[0], 'valid samples')
print(test_images.shape[0], 'test samples')
# 我们的模型使用categorical_crossentropy作为损失函数，因此需要根据类别数量nb_classes将
# 类别标签进行one-hot编码使其向量化，在这里我们的类别只有两种，经过转化后标签数据变为二维
train_labels = np_utils.to_categorical(train_labels, nb_classes)
valid_labels = np_utils.to_categorical(valid_labels, nb_classes)
test_labels = np_utils.to_categorical(test_labels, nb_classes)
# 像素数据浮点化以便归一化 目的是提升网络收敛速度，减少训练时间，同时适应值域在（0，1）之间的激活函数，增大区分度。
#归一化有一个特别重要的原因是确保特征值权重一致
train_images = train_images.astype('float32')
valid_images = valid_images.astype('float32')
test_images = test_images.astype('float32')
# 将其归一化,图像的各像素值归一化到0~1区间
train_images /= 255
valid_images /= 255
test_images /= 255
self.train_images = train_images
self.valid_images = valid_images
self.test_images = test_images
self.train_labels = train_labels
self.valid_labels = valid_labels
self.test_labels = test_labels
2、训练完后的模型保存
3、保存路径为model文件下
"""
if __name__ == '__main__':
# 读取数据
dataset = Dataset(r'E:\sign_system\face_asia_500_crop')
# 读取路径
path_name = r'E:\sign_system\face_asia_500_crop'
dataset.load()
model = Model_train()
# 训练模型
_, _, num_classes = load_dataset(path_name)
print(num_classes)
print('\nTrain_Starting--------------------------------------------------')
model.Squeezenet(num_classes, dataset)
model.train(dataset)
print('Model Saved.')
model.save_model(file_path = 'E:/sign_system/execute_system/haar_extract/squeezenetface4.h5')
# 评估模型
# model = Model_train()
# print('\nTesting---------------------------------------------------------')
# model.load_model(file_path = 'C:\Users\Administrator\Desktop\FaceRecognition_Version3\model\squeezenet.model.h5')
def load(self,
img_rows=IMAGE_SIZE,
img_cols=IMAGE_SIZE,
img_channels=3,
nb_classes=2):
#加载数据集到内存
images, labels = load_dataset(self.path_name)
#/////第一步，交叉验证。划分训练集和验证集，以及测试集
train_images, valid_images, train_labels, valid_labels = train_test_split(
images, labels, test_size=0.3, random_state=random.randint(0, 100))
_, test_images, _, test_labels = train_test_split(
images, labels, test_size=0.5, random_state=random.randint(0, 100))
#当前的维度顺序如果是‘th’，则输入图片数据时的顺序为：channels,rows,
#cols,否则是：rows,cols,channels
#根据keras库要的维度顺序重组训练数据集
#print(K.image_data_format())
#/////第二步，根据后端系统，重新调整数组的维度
if K.image_data_format() == 'channels_first':
train_images = train_images.reshape(train_images.shape[0],
img_channels, img_rows,
img_cols)
valid_images = valid_images.reshape(valid_images.shape[0],
img_channels, img_rows,
img_cols)
test_images = test_images.reshape(test_images.shape[0],
img_channels, img_rows, img_cols)
self.input_shape = (img_channels, img_rows, img_cols)
else:
train_images = train_images.reshape(train_images.shape[0],
img_rows, img_cols,
img_channels)
valid_images = valid_images.reshape(valid_images.shape[0],
img_rows, img_cols,
img_channels)
test_images = test_images.reshape(test_images.shape[0], img_rows,
img_cols, img_channels)
self.input_shape = (img_rows, img_cols, img_channels)
#输出训练集、验证集 、测试集的数量
print(train_images.shape[0], 'train samples')
print(valid_images.shape[0], 'valid samples')
print(test_images.shape[0], 'test samples')
#模型使用categorical_crossentropy作为损失函数，因此需要根据类别
#数量nb_classes将类别标签进行one-hot编码使其向量化，
#经过转化后标签数据变为二维
#/////第三步，一位有效编码。针对标签。
train_labels = np_utils.to_categorical(train_labels, nb_classes)
valid_labels = np_utils.to_categorical(valid_labels, nb_classes)
test_labels = np_utils.to_categorical(test_labels, nb_classes)
#像素数据浮点化以便归一化
#/////第四步，先浮点后归一化。目的，提升网络收敛速度，减少训练
#/////时间，同事适应值在（0，1）之间的激活函数，增大区分度。
#归一化有一重要特性，就是保持特征值权重一致。
train_images = train_images.astype('float32')
valid_images = valid_images.astype('float32')
test_images = test_images.astype('float32')
#将其归一化，图像各像素归一化到0-1区间
train_images /= 255
valid_images /= 255
test_images /= 255
self.train_images = train_images
self.valid_images = valid_images
self.test_images = test_images
self.train_labels = train_labels
self.valid_labels = valid_labels
self.test_labels = test_labels
def load(self, img_rows=IMAGE_SIZE, img_cols=IMAGE_SIZE, img_channels=3):
# 加载数据集到内存
images, labels = load_dataset(self.path_name)
train_images, valid_images, train_labels, valid_labels = \
train_test_split(images, labels, test_size=0.1,random_state=random.randint(0, 100))
#print(train_labels)
valid_images, test_images, valid_labels, test_labels = \
train_test_split(valid_images, valid_labels, test_size=0.5,random_state=random.randint(0, 100))
# 当前的维度顺序如果为'th'，则输入图片数据时的顺序为：channels,rows,cols，否则:rows,cols,channels
# 这部分代码就是根据keras库要求的维度顺序重组训练数据集
if K.image_dim_ordering() == 'th':
train_images = train_images.reshape(train_images.shape[0],
img_channels, img_rows,
img_cols)
valid_images = valid_images.reshape(valid_images.shape[0],
img_channels, img_rows,
img_cols)
test_images = test_images.reshape(test_images.shape[0],
img_channels, img_rows, img_cols)
self.input_shape = (img_channels, img_rows, img_cols)
else:
train_images = train_images.reshape(train_images.shape[0],
img_rows, img_cols,
img_channels)
valid_images = valid_images.reshape(valid_images.shape[0],
img_rows, img_cols,
img_channels)
test_images = test_images.reshape(test_images.shape[0], img_rows,
img_cols, img_channels)
self.input_shape = (img_rows, img_cols, img_channels)
# 输出训练集、验证集、测试集的数量
print(train_images.shape[0], 'train samples')
print(valid_images.shape[0], 'valid samples')
print(test_images.shape[0], 'test samples')
# 我们的模型使用categorical_crossentropy（交叉熵损失函数）作为损失函数，
# 该函数要求标签集必须采用独热编码的形式，
# 因此需要根据类别数量nb_classes将
# 类别标签进行one-hot编码使其向量化
train_labels = np_utils.to_categorical(train_labels)
valid_labels = np_utils.to_categorical(valid_labels)
test_labels = np_utils.to_categorical(test_labels)
print(len(train_labels[1, ]))
# 像素数据浮点化以便归一化
train_images = train_images.astype('float32')
valid_images = valid_images.astype('float32')
test_images = test_images.astype('float32')
# 将其归一化,图像的各像素值归一化到0~1区间
# 目的是提升网络收敛速度，减少训练时间，同时适应值域在(0,1)之间的激活函数，
# 此外，确保特征值权重一致，避免大值对误差值有较大影响
train_images /= 255
valid_images /= 255
test_images /= 255
self.train_images = train_images
self.valid_images = valid_images
self.test_images = test_images
self.train_labels = train_labels
self.valid_labels = valid_labels
self.test_labels = test_labels
self.nb_classes = len(train_labels[0])
def load(self,
img_rows=IMAGE_SIZE,
img_cols=IMAGE_SIZE,
img_channels=3,
nb_classes=0):
# 載入數據到記憶體
images, labels = load_dataset(self.path_name)
train_images, valid_images, train_labels, valid_labels = train_test_split(
images, labels, test_size=0.3, random_state=random.randint(0, 100))
_, test_images, _, test_labels = train_test_split(
images, labels, test_size=0.5, random_state=random.randint(0, 100))
# 如果現在的維度是 th 則輸入的順序為 channels,rows,cols 否則是 rows,cols,channels
# 根據 keras 要求的維度順序訓練模組
if K.image_dim_ordering() == 'th':
train_images = train_images.reshape(train_images.shape[0],
img_channels, img_rows,
img_cols)
valid_images = valid_images.reshape(valid_images.shape[0],
img_channels, img_rows,
img_cols)
test_images = test_images.reshape(test_images.shape[0],
img_channels, img_rows, img_cols)
self.input_shape = (img_channels, img_rows, img_cols)
else:
train_images = train_images.reshape(train_images.shape[0],
img_rows, img_cols,
img_channels)
valid_images = valid_images.reshape(valid_images.shape[0],
img_rows, img_cols,
img_channels)
test_images = test_images.reshape(test_images.shape[0], img_rows,
img_cols, img_channels)
self.input_shape = (img_rows, img_cols, img_channels)
# 印出訓練圖片、驗證圖片、測試圖片的數量
print(train_images.shape[0], 'train samples')
print(valid_images.shape[0], 'valid samples')
print(test_images.shape[0], 'test samples')
# 此模型用 categorical_crossentropy 作為損失函數，將圖片根據要訓練的人數向量化
train_labels = np_utils.to_categorical(train_labels, nb_classes)
valid_labels = np_utils.to_categorical(valid_labels, nb_classes)
test_labels = np_utils.to_categorical(test_labels, nb_classes)
# 將數值標準化
train_images = train_images.astype('float32')
valid_images = valid_images.astype('float32')
test_images = test_images.astype('float32')
# 將數值標準化至 0-1 之間
train_images /= 255
valid_images /= 255
test_images /= 255
self.train_images = train_images
self.valid_images = valid_images
self.test_images = test_images
self.train_labels = train_labels
self.valid_labels = valid_labels
self.test_labels = test_labels
def build_model(self, dataset, path_name):
# 确定目标的类别
_, _, nb_classes = load_dataset(path_name)
# 构建一个空的网络模型，它是一个线性堆叠模型，各神经网络层会被顺序添加，专业名称为序贯模型或线性堆叠模型
self.model = Sequential()
# 以下代码将顺序添加CNN网络需要的各层，一个add就是一个网络层
#1 2维卷积层1，卷积核为32 × 3 × 3,边界为"same"就是保持卷积后图像大小不变，就是需要padding填充
# 输入图形一般就是数据集的形状，然后输出为64 × 64 × 32，参数为（3 × 3 × 3 + 1） × 32 = 896
self.model.add(
Conv2D(32, (3, 3), padding='same',
input_shape=dataset.input_shape))
#2 激活函数层，输出为64 × 64 × 32
self.model.add(Activation('relu'))
#3 2维卷积层2，卷积核为32 × 3 × 3,输出为62 × 62 × 32，参数为9248
self.model.add(Conv2D(32, (3, 3)))
#4 激活函数层，输出为62 × 62 × 32
self.model.add(Activation('relu'))
#5 池化层1，池化区域为2 × 2，输出为31 × 31 × 32
self.model.add(MaxPooling2D(pool_size=(2, 2)))
#6 Dropout层1，dropout率为0.25，输出为31 × 31 × 32
self.model.add(Dropout(0.25))
#7 2维卷积层3，卷积核为64 × 3 × 3，边界为"same"就是保持卷积后图像大小不变，就是需要padding填充
# 输出为31 × 31 × 64，参数为18496
self.model.add(Conv2D(64, (3, 3), padding='same'))
#8 激活函数层，输出为31 × 31 × 64
self.model.add(Activation('relu'))
#9 2维卷积层4，卷积核为32 × 3 × 3,输出为29 × 29 × 64，参数为36928
self.model.add(Conv2D(64, (3, 3)))
#10 激活函数层，输出为输出为29 × 29 × 64
self.model.add(Activation('relu'))
#11 池化层2，池化区域为2 × 2，输出为14 × 14 × 64
self.model.add(MaxPooling2D(pool_size=(2, 2)))
#12 Dropout层2，dropout率为0.25，输出为14 × 14 × 64
self.model.add(Dropout(0.25))
#13 Flatten层，输出14 × 14 × 64=12455
self.model.add(Flatten())
#14 Dense层1,又被称作全连接层，输出为512，参数为6423040
self.model.add(Dense(512))
#15 激活函数层，输出为512
self.model.add(Activation('relu'))
#16 Dropout层3，dropout率设置为0.5
self.model.add(Dropout(0.5))
#17 Dense层2，输出属于各类别概率，参数为1026
self.model.add(Dense(nb_classes))
#18 分类层，输出最终结果
self.model.add(Activation('softmax'))
#输出模型概况
self.model.summary()
return nb_classes
def load(self, img_rows=224, img_cols=224, img_channels=3, nb_classes=111):
#加载数据集到内存
images, labels = load_dataset(self.path_name)
test_images = images[range(2, 777, 7)]
valid_images = images[range(1, 777, 7)]
train_images = images[list(range(0, 777, 7)) + list(range(3, 777, 7)) +
list(range(4, 777, 7)) + list(range(5, 777, 7)) +
list(range(6, 777, 7))]
test_labels = labels[2:777:7]
valid_labels = labels[1:777:7]
train_labels = labels[0:777:7] + labels[3:777:7] + labels[
4:777:7] + labels[5:777:7] + labels[6:777:7]
#train_images, valid_images, train_labels, valid_labels = train_test_split(images, labels, test_size = 0.3, random_state = random.randint(0, 100))
#_, test_images, _, test_labels = train_test_split(images, labels, test_size = 0.5, random_state = random.randint(0, 100))
#当前的维度顺序如果为'th'，则输入图片数据时的顺序为：channels,rows,cols，否则:rows,cols,channels
#这部分代码就是根据keras库要求的维度顺序重组训练数据集
if K.image_dim_ordering() == 'th':
train_images = train_images.reshape(train_images.shape[0],
img_channels, img_rows,
img_cols)
valid_images = valid_images.reshape(valid_images.shape[0],
img_channels, img_rows,
img_cols)
test_images = test_images.reshape(test_images.shape[0],
img_channels, img_rows, img_cols)
self.input_shape = (img_channels, img_rows, img_cols)
else:
train_images = train_images.reshape(train_images.shape[0],
img_rows, img_cols,
img_channels)
valid_images = valid_images.reshape(valid_images.shape[0],
img_rows, img_cols,
img_channels)
test_images = test_images.reshape(test_images.shape[0], img_rows,
img_cols, img_channels)
self.input_shape = (img_rows, img_cols, img_channels)
#输出训练集、验证集、测试集的数量
print(train_images.shape[0], 'train samples')
print(valid_images.shape[0], 'valid samples')
print(test_images.shape[0], 'test samples')
#我们的模型使用categorical_crossentropy作为损失函数，因此需要根据类别数量nb_classes将
#类别标签进行one-hot编码使其向量化，在这里我们的类别只有两种，经过转化后标签数据变为二维
# train_labels = np_utils.to_categorical(train_labels, nb_classes)
# valid_labels = np_utils.to_categorical(valid_labels, nb_classes)
# test_labels = np_utils.to_categorical(test_labels, nb_classes)
#像素数据浮点化以便归一化
train_images = train_images.astype('float32')
valid_images = valid_images.astype('float32')
test_images = test_images.astype('float32')
#将其归一化,图像的各像素值归一化到0~1区间
train_images /= 255
valid_images /= 255
test_images /= 255
lb = LabelBinarizer().fit(np.array(range(0, nb_classes)))
self.train_images = train_images
self.valid_images = valid_images
self.test_images = test_images
self.train_labels = lb.transform(train_labels)
self.valid_labels = lb.transform(valid_labels)
self.test_labels = lb.transform(test_labels)
def load(self,
img_rows=IMAGE_SIZE,
img_cols=IMAGE_SIZE,
img_channels=3,
nb_classes=2):
#加载数据到内存
images, labels = load_dataset(self.path_name)
train_images,valid_images,train_labels,valid_labels = \
train_test_split(images,labels,test_size=0.3,random_state=random.randint(0,100))
_, test_images, _, test_labels = train_test_split(
images, labels, test_size=0.5, random_state=random.randint(0, 100))
# 当前的维度顺序如果为'th'，则输入图片数据时的顺序为：channels,rows,cols，否则:rows,cols,channels
# 这部分代码就是根据keras库要求的维度顺序重组训练数据集
if k.image_dim_ordering() == 'th':
train_images = train_images.reshape(train_images.shape[0],
img_channels, img_rows,
img_cols)
valid_images = valid_images.reshape(valid_images.shape[0],
img_channels, img_rows,
img_cols)
test_images = test_images.reshape(test_images.shape[0],
img_channels, img_rows, img_cols)
self.input_shape = (img_channels, img_rows, img_cols)
else:
train_images = train_images.reshape(train_images.shape[0],
img_rows, img_cols,
img_channels)
valid_images = valid_images.reshape(valid_images.shape[0],
img_rows, img_cols,
img_channels)
test_images = test_images.reshape(test_images.shape[0], img_rows,
img_cols, img_channels)
self.input_shape = (img_rows, img_cols, img_channels)
# 我们的模型使用categorical_crossentropy作为损失函数，因此需要根据类别数量nb_classes将
# 类别标签进行one-hot编码使其向量化，在这里我们的类别只有两种，经过转化后标签数据变为二维
train_labels = np_utils.to_categorical(train_labels, nb_classes)
valid_labels = np_utils.to_categorical(valid_labels, nb_classes)
test_labels = np_utils.to_categorical(test_labels, nb_classes)
# 像素数据浮点化以便归一化
train_images = train_images.astype('float32')
valid_images = valid_images.astype('float32')
test_images = test_images.astype('float32')
#将其归一化,图像的各像素值归一化到0~1区间
train_images /= 255
valid_images /= 255
test_images /= 255
#数据集先浮点后归一化的目的是提升网络收敛速度，减少训练时间，同时适应值域在（0,1）之间的激活函数，
# 增大区分度。其实归一化有一个特别重要的原因是确保特征值权重一致。举个例子，我们使用mse这样的均方误差函数时，
# 大的特征数值比如(5000-1000)2与小的特征值(3-1)2相加再求平均得到的误差值，显然大值对误差值的影响最大，但大部分情况下，
# 特征值的权重应该是一样的，只是因为单位不同才导致数值相差甚大。因此，我们提前对特征数据做归一化处理，以解决此类问题。
self.train_images = train_images
self.train_labels = train_labels
self.valid_images = valid_images
self.valid_labels = valid_labels
self.test_images = test_images
self.test_labels = test_labels
# -*- coding: utf-8 -*-
from keras.models import Sequential
from keras.layers import Dense, Conv2D, MaxPooling2D, Dropout, Flatten
from keras.models import load_model
from load_face_dataset import load_dataset
from keras.optimizers import SGD
from keras.preprocessing.image import ImageDataGenerator
import numpy
# fix random seed for reproducibility
seed = 7
numpy.random.seed(seed)
# 导入皮马印第安人糖尿病数据集. 它描述了病人医疗记录和他们是否在五年内发病。
images, labels = load_dataset('../faces')
# 8个输入变量 1个输出变量，最后一列为输出 1为出现糖尿病，0为未出现
X = images # 输入
Y = labels # 输出
# model = Sequential()
# model.add(Conv2D(64, (3, 3), padding='same', input_shape=(64, 64, 3), activation='relu'))
# model.add(MaxPooling2D(pool_size=(2, 2)))
# model.add(Conv2D(32, (3, 3), activation='relu'))
# model.add(MaxPooling2D(pool_size=(2, 2)))
# model.add(Dropout(0.25))
# model.add(Conv2D(32, (3, 3), activation='relu'))
# model.add(MaxPooling2D(pool_size=(2, 2)))
# model.add(Dropout(0.25))
# model.add(Flatten()) # 13 Flatten层
# model.add(Dense(128, activation='relu')) # 14 Dense层,又被称作全连接层
from keras.models import load_model
from load_face_dataset import load_dataset, resize_image, IMAGE_SIZE
model = load_model("model_weight.h5")
print(model)
images, _ = load_dataset('D:/tmp/output/')
image = images.reshape(images.shape[0], 64, 64, 3)
result = model.predict_classes(image)
print('Predicted:', result)
