代码来自this github repo . repo为模型提供代码，我们的目标是训练模型，基于类似于OpenFace模型实现的三元组丢失函数生成128个嵌入 .
型号代码如下所述 . 仅在使用opencv调整图像大小后才提供输入
def createmodel（）：
myInput = Input(shape=(96,96,3))
x = ZeroPadding2D(padding=(3, 3), input_shape=(96,96,3))(myInput)
x = Conv2D(64, (7, 7), strides=(2, 2), name='conv1')(x)
x = BatchNormalization(axis=3, epsilon=0.00001, name='bn1')(x)
x = Activation('relu')(x)
x = ZeroPadding2D(padding=(1, 1))(x)
x = MaxPooling2D(pool_size=3, strides=2)(x)
x = Lambda(LRN2D, name='lrn_1')(x)
x = Conv2D(64, (1, 1), name='conv2')(x)
x = BatchNormalization(axis=3, epsilon=0.00001, name='bn2')(x)
x = Activation('relu')(x)
x = ZeroPadding2D(padding=(1, 1))(x)
x = Conv2D(192, (3, 3), name='conv3')(x)
x = BatchNormalization(axis=3, epsilon=0.00001, name='bn3')(x)
x = Activation('relu')(x)
x = Lambda(LRN2D, name='lrn_2')(x)
x = ZeroPadding2D(padding=(1, 1))(x)
x = MaxPooling2D(pool_size=3, strides=2)(x)
# Inception3a
inception_3a_3x3 = Conv2D(96, (1, 1), name='inception_3a_3x3_conv1')(x)
inception_3a_3x3 = BatchNormalization(axis=3, epsilon=0.00001, name='inception_3a_3x3_bn1')(inception_3a_3x3)
inception_3a_3x3 = Activation('relu')(inception_3a_3x3)
inception_3a_3x3 = ZeroPadding2D(padding=(1, 1))(inception_3a_3x3)
inception_3a_3x3 = Conv2D(128, (3, 3), name='inception_3a_3x3_conv2')(inception_3a_3x3)
inception_3a_3x3 = BatchNormalization(axis=3, epsilon=0.00001, name='inception_3a_3x3_bn2')(inception_3a_3x3)
inception_3a_3x3 = Activation('relu')(inception_3a_3x3)
inception_3a_5x5 = Conv2D(16, (1, 1), name='inception_3a_5x5_conv1')(x)
inception_3a_5x5 = BatchNormalization(axis=3, epsilon=0.00001, name='inception_3a_5x5_bn1')(inception_3a_5x5)
inception_3a_5x5 = Activation('relu')(inception_3a_5x5)
inception_3a_5x5 = ZeroPadding2D(padding=(2, 2))(inception_3a_5x5)
inception_3a_5x5 = Conv2D(32, (5, 5), name='inception_3a_5x5_conv2')(inception_3a_5x5)
inception_3a_5x5 = BatchNormalization(axis=3, epsilon=0.00001, name='inception_3a_5x5_bn2')(inception_3a_5x5)
inception_3a_5x5 = Activation('relu')(inception_3a_5x5)
inception_3a_pool = MaxPooling2D(pool_size=3, strides=2)(x)
inception_3a_pool = Conv2D(32, (1, 1), name='inception_3a_pool_conv')(inception_3a_pool)
inception_3a_pool = BatchNormalization(axis=3, epsilon=0.00001, name='inception_3a_pool_bn')(inception_3a_pool)
inception_3a_pool = Activation('relu')(inception_3a_pool)
inception_3a_pool = ZeroPadding2D(padding=((3, 4), (3, 4)))(inception_3a_pool)
inception_3a_1x1 = Conv2D(64, (1, 1), name='inception_3a_1x1_conv')(x)
inception_3a_1x1 = BatchNormalization(axis=3, epsilon=0.00001, name='inception_3a_1x1_bn')(inception_3a_1x1)
inception_3a_1x1 = Activation('relu')(inception_3a_1x1)
inception_3a = concatenate([inception_3a_3x3, inception_3a_5x5, inception_3a_pool, inception_3a_1x1], axis=3)
# Inception3b
inception_3b_3x3 = Conv2D(96, (1, 1), name='inception_3b_3x3_conv1')(inception_3a)
inception_3b_3x3 = BatchNormalization(axis=3, epsilon=0.00001, name='inception_3b_3x3_bn1')(inception_3b_3x3)
inception_3b_3x3 = Activation('relu')(inception_3b_3x3)
inception_3b_3x3 = ZeroPadding2D(padding=(1, 1))(inception_3b_3x3)
inception_3b_3x3 = Conv2D(128, (3, 3), name='inception_3b_3x3_conv2')(inception_3b_3x3)
inception_3b_3x3 = BatchNormalization(axis=3, epsilon=0.00001, name='inception_3b_3x3_bn2')(inception_3b_3x3)
inception_3b_3x3 = Activation('relu')(inception_3b_3x3)
inception_3b_5x5 = Conv2D(32, (1, 1), name='inception_3b_5x5_conv1')(inception_3a)
inception_3b_5x5 = BatchNormalization(axis=3, epsilon=0.00001, name='inception_3b_5x5_bn1')(inception_3b_5x5)
inception_3b_5x5 = Activation('relu')(inception_3b_5x5)
inception_3b_5x5 = ZeroPadding2D(padding=(2, 2))(inception_3b_5x5)
inception_3b_5x5 = Conv2D(64, (5, 5), name='inception_3b_5x5_conv2')(inception_3b_5x5)
inception_3b_5x5 = BatchNormalization(axis=3, epsilon=0.00001, name='inception_3b_5x5_bn2')(inception_3b_5x5)
inception_3b_5x5 = Activation('relu')(inception_3b_5x5)
inception_3b_pool = Lambda(lambda x: x**2, name='power2_3b')(inception_3a)
inception_3b_pool = AveragePooling2D(pool_size=(3, 3), strides=(3, 3))(inception_3b_pool)
inception_3b_pool = Lambda(lambda x: x*9, name='mult9_3b')(inception_3b_pool)
inception_3b_pool = Lambda(lambda x: K.sqrt(x), name='sqrt_3b')(inception_3b_pool)
inception_3b_pool = Conv2D(64, (1, 1), name='inception_3b_pool_conv')(inception_3b_pool)
inception_3b_pool = BatchNormalization(axis=3, epsilon=0.00001, name='inception_3b_pool_bn')(inception_3b_pool)
inception_3b_pool = Activation('relu')(inception_3b_pool)
inception_3b_pool = ZeroPadding2D(padding=(4, 4))(inception_3b_pool)
inception_3b_1x1 = Conv2D(64, (1, 1), name='inception_3b_1x1_conv')(inception_3a)
inception_3b_1x1 = BatchNormalization(axis=3, epsilon=0.00001, name='inception_3b_1x1_bn')(inception_3b_1x1)
inception_3b_1x1 = Activation('relu')(inception_3b_1x1)
inception_3b = concatenate([inception_3b_3x3, inception_3b_5x5, inception_3b_pool, inception_3b_1x1], axis=3)
# Inception3c
inception_3c_3x3 = utils.conv2d_bn(inception_3b,
layer='inception_3c_3x3',
cv1_out=128,
cv1_filter=(1, 1),
cv2_out=256,
cv2_filter=(3, 3),
cv2_strides=(2, 2),
padding=(1, 1))
inception_3c_5x5 = utils.conv2d_bn(inception_3b,
layer='inception_3c_5x5',
cv1_out=32,
cv1_filter=(1, 1),
cv2_out=64,
cv2_filter=(5, 5),
cv2_strides=(2, 2),
padding=(2, 2))
inception_3c_pool = MaxPooling2D(pool_size=3, strides=2)(inception_3b)
inception_3c_pool = ZeroPadding2D(padding=((0, 1), (0, 1)))(inception_3c_pool)
inception_3c = concatenate([inception_3c_3x3, inception_3c_5x5, inception_3c_pool], axis=3)
#inception 4a
inception_4a_3x3 = utils.conv2d_bn(inception_3c,
layer='inception_4a_3x3',
cv1_out=96,
cv1_filter=(1, 1),
cv2_out=192,
cv2_filter=(3, 3),
cv2_strides=(1, 1),
padding=(1, 1))
inception_4a_5x5 = utils.conv2d_bn(inception_3c,
layer='inception_4a_5x5',
cv1_out=32,
cv1_filter=(1, 1),
cv2_out=64,
cv2_filter=(5, 5),
cv2_strides=(1, 1),
padding=(2, 2))
inception_4a_pool = Lambda(lambda x: x**2, name='power2_4a')(inception_3c)
inception_4a_pool = AveragePooling2D(pool_size=(3, 3), strides=(3, 3))(inception_4a_pool)
inception_4a_pool = Lambda(lambda x: x*9, name='mult9_4a')(inception_4a_pool)
inception_4a_pool = Lambda(lambda x: K.sqrt(x), name='sqrt_4a')(inception_4a_pool)
inception_4a_pool = utils.conv2d_bn(inception_4a_pool,
layer='inception_4a_pool',
cv1_out=128,
cv1_filter=(1, 1),
padding=(2, 2))
inception_4a_1x1 = utils.conv2d_bn(inception_3c,
layer='inception_4a_1x1',
cv1_out=256,
cv1_filter=(1, 1))
inception_4a = concatenate([inception_4a_3x3, inception_4a_5x5, inception_4a_pool, inception_4a_1x1], axis=3)
#inception4e
inception_4e_3x3 = utils.conv2d_bn(inception_4a,
layer='inception_4e_3x3',
cv1_out=160,
cv1_filter=(1, 1),
cv2_out=256,
cv2_filter=(3, 3),
cv2_strides=(2, 2),
padding=(1, 1))
inception_4e_5x5 = utils.conv2d_bn(inception_4a,
layer='inception_4e_5x5',
cv1_out=64,
cv1_filter=(1, 1),
cv2_out=128,
cv2_filter=(5, 5),
cv2_strides=(2, 2),
padding=(2, 2))
inception_4e_pool = MaxPooling2D(pool_size=3, strides=2)(inception_4a)
inception_4e_pool = ZeroPadding2D(padding=((0, 1), (0, 1)))(inception_4e_pool)
inception_4e = concatenate([inception_4e_3x3, inception_4e_5x5, inception_4e_pool], axis=3)
#inception5a
inception_5a_3x3 = utils.conv2d_bn(inception_4e,
layer='inception_5a_3x3',
cv1_out=96,
cv1_filter=(1, 1),
cv2_out=384,
cv2_filter=(3, 3),
cv2_strides=(1, 1),
padding=(1, 1))
inception_5a_pool = Lambda(lambda x: x**2, name='power2_5a')(inception_4e)
inception_5a_pool = AveragePooling2D(pool_size=(3, 3), strides=(3, 3))(inception_5a_pool)
inception_5a_pool = Lambda(lambda x: x*9, name='mult9_5a')(inception_5a_pool)
inception_5a_pool = Lambda(lambda x: K.sqrt(x), name='sqrt_5a')(inception_5a_pool)
inception_5a_pool = utils.conv2d_bn(inception_5a_pool,
layer='inception_5a_pool',
cv1_out=96,
cv1_filter=(1, 1),
padding=(1, 1))
inception_5a_1x1 = utils.conv2d_bn(inception_4e,
layer='inception_5a_1x1',
cv1_out=256,
cv1_filter=(1, 1))
inception_5a = concatenate([inception_5a_3x3, inception_5a_pool, inception_5a_1x1], axis=3)
#inception_5b
inception_5b_3x3 = utils.conv2d_bn(inception_5a,
layer='inception_5b_3x3',
cv1_out=96,
cv1_filter=(1, 1),
cv2_out=384,
cv2_filter=(3, 3),
cv2_strides=(1, 1),
padding=(1, 1))
inception_5b_pool = MaxPooling2D(pool_size=3, strides=2)(inception_5a)
inception_5b_pool = utils.conv2d_bn(inception_5b_pool,
layer='inception_5b_pool',
cv1_out=96,
cv1_filter=(1, 1))
inception_5b_pool = ZeroPadding2D(padding=(1, 1))(inception_5b_pool)
inception_5b_1x1 = utils.conv2d_bn(inception_5a,
layer='inception_5b_1x1',
cv1_out=256,
cv1_filter=(1, 1))
inception_5b = concatenate([inception_5b_3x3, inception_5b_pool, inception_5b_1x1], axis=3)
av_pool = AveragePooling2D(pool_size=(3, 3), strides=(1, 1))(inception_5b)
reshape_layer = Flatten()(av_pool)
dense_layer = Dense(128, name='dense_layer')(reshape_layer)
norm_layer = Lambda(lambda x: K.l2_normalize(x, axis=1), name='norm_layer')(dense_layer)
# Final Model
return Model(inputs=[myInput], outputs=norm_layer)
from keras import backend as K
from keras.models import Model
from keras.layers import Input, Layer
# Input for anchor, positive and negative images
in_a = Input(shape=(96, 96, 3))
in_p = Input(shape=(96, 96, 3))
in_n = Input(shape=(96, 96, 3))
# Output for anchor, positive and negative embedding vectors
# The nn4_small model instance is shared (Siamese network)
emb_a = nn4_small2(in_a)
emb_p = nn4_small2(in_p)
emb_n = nn4_small2(in_n)
class TripletLossLayer(Layer):
def __init__(self, alpha, **kwargs):
self.alpha = alpha
super(TripletLossLayer, self).__init__(**kwargs)
def triplet_loss(self, inputs):
a, p, n = inputs
p_dist = K.sum(K.square(a-p), axis=-1)
n_dist = K.sum(K.square(a-n), axis=-1)
return K.sum(K.maximum(p_dist - n_dist + self.alpha, 0), axis=0)
def call(self, inputs):
loss = self.triplet_loss(inputs)
self.add_loss(loss)
return loss
# Layer that computes the triplet loss from anchor, positive and negative embedding vectors
triplet_loss_layer = TripletLossLayer(alpha=0.2, name='triplet_loss_layer')([emb_a, emb_p, emb_n])
# Model that can be trained with anchor, positive negative images
nn4_small2_train = Model([in_a, in_p, in_n], triplet_loss_layer)
使用发电机训练模型 . 下面提到代码，首先根据参数people_per_batch，images_per_person对一组图像进行采样 . 接下来，使用采样图像创建三元组 . triplet_path_to_numpy_representation（）用于使用opencv从路径表示转换为numpy数组图像
def triplet_path_to_numpy_representation(triplet_array):
triplet_numpy_array = []
for triplet in triplet_array:
# print(triplet)
triplet_numpy = []
for individual_image in triplet:
img=cv2.imread(individual_image)
img=cv2.resize(img,(96,96))
triplet_numpy.append(img)
triplet_numpy_array.append(triplet_numpy)
return triplet_numpy_array
def triplet_generator(model,people_per_batch, images_per_person,embedding_size,alpha):
''' Dummy triplet generator for API usage demo only.
Will be replaced by a version that uses real image data later.
:return: a batch of (anchor, positive, negative) triplets
'''
dataset=get_dataset("test")
nrof_examples = people_per_batch * images_per_person
while True:
image_paths, num_per_class = sample_people(dataset, people_per_batch, images_per_person)
labels_array = np.arange(nrof_examples)
image_paths_array = np.reshape(np.expand_dims(np.array(image_paths),1), (-1,3))
# sess.run(enqueue_op, {image_paths_placeholder: image_paths_array, labels_placeholder: labels_array})
emb_array = np.zeros((nrof_examples, embedding_size))
# nrof_batches = int(np.ceil(nrof_examples / batch_size))
# for i in range(nrof_batches):
# batch_size = min(nrof_examples-i*batch_size, batch_size)
triplet_numpy_array=triplet_path_to_numpy_representation(image_paths_array)
# namelist.append(image_paths_array)
# embeddinglist.append(triplet_numpy_array)
emb=nn4_small2.predict(np.reshape(np.array(triplet_numpy_array),(-1,96,96,3)))
# xlist.append(emb)
# emb=model.predict(np.reshape(np.array(triplet_numpy_array),(-1,200,180,3)))
emb_array[labels_array,:] = emb
triplets, nrof_random_negs, nrof_triplets = select_triplets(emb_array, num_per_class,
image_paths, people_per_batch, alpha)
labels_array = np.reshape(np.arange(len(triplets)*3),(-1,3))
triplet_paths_array = np.reshape(np.expand_dims(np.array(triplets),1), (-1,3))
triplet_paths_array=np.array(triplet_path_to_numpy_representation(triplet_paths_array))
temp=np.hsplit(triplet_paths_array,3)
a_batch = np.reshape(temp[0],(-1,96,96,3))
p_batch = np.reshape(temp[1],(-1,96,96,3))
n_batch = np.reshape(temp[2],(-1,96,96,3))
# a_batch = np.reshape(temp[0],(-1,200,180,3))
# p_batch = np.reshape(temp[1],(-1,200,180,3))
# n_batch = np.reshape(temp[2],(-1,200,180,3))
yield [a_batch , p_batch, n_batch], None
在使用fit_generator函数进行训练时调用的triplet_generator函数中的行中出现问题：
generator = triplet_generator(5,3,128,0.2)
nn4_small2_train.compile(loss=, optimizer='adam')
nn4_small2_train.fit_generator(generator, epochs=1, steps_per_epoch=1)
emb = nn4_small2.predict（np.reshape（np.array（triplet_numpy_array），（ - 1,96,96,3）））
甚至在模型上进行单个历元训练之后，调用nn4_small2.predict（img），其中img是形状图像（96,96,3），输出是嵌入形状（1,128）但是所有值都是Nan .
但是在训练之前调用预测会生成没有任何Nan值的嵌入 .
任何帮助，将不胜感激 .