相似图像搜索从训练到服务全过程

def add_new_last_layer(base_model, nb_classes):

    '''
    添加最后的层
    :param base_model: 预训练模型
    :param nb_classes: 分类数量
    :return: 新的 model
    '''
    x = base_model.output
    x = GlobalAveragePooling2D()(x)
    x = Dense(128, activation='relu')(x) #输出的特征维度 88
    predictions = Dense(nb_classes, activation='softmax')(x)
    model = Model(input=base_model.input, output=predictions)
    return model

def freeze_base_layer(model, base_model):

        for layer in base_model.layers:

        layer.trainable = False

model.compile(optimizer='rmsprop', loss='categorical_crossentropy', metrics= ['accuracy'])

# optimizer: 优化器

# loss: 损失函数，多类的对数损失需要将分类标签转换为（将标签转化为形如(nb_samples, nb_classes)的二值序列）

# metrics: 列表，包含评估模型在训练和测试时的网络性能的指标准备训练数据。

#数据准备
IM_WIDTH, IM_HEIGHT = 224,224
train_dir = './refine_img_data/train'
val_dir = './refine_img_data/test'
nb_classes = 5
np_epoch = 3
batch_size = 16
nb_train_samples = get_nb_files(train_dir)
nb_classes = len(glob.glob(train_dir + '/*'))
nb_val_samples = get_nb_files(val_dir)

# 根据现有数据，设置新数据生成参数
train_datagen = ImageDataGenerator(
preprocessing_function=preprocess_input,
rotation_range=30,
width_shift_range=0.2,
height_shift_range=0.2,
shear_range=0.2,
zoom_range=0.2,
horizontal_flip=True
)

test_datagen = ImageDataGenerator(
preprocessing_function=preprocess_input,
rotation_range=30,
width_shift_range=0.2,
height_shift_range=0.2,
shear_range=0.2,
zoom_range=0.2,
horizontal_flip=True
)

# 从文件夹获取数据
train_generator = train_datagen.flow_from_directory(
train_dir,
target_size=(IM_WIDTH, IM_HEIGHT),
batch_size=batch_size,
class_mode='categorical'
)

validation_generator = test_datagen.flow_from_directory(
val_dir,
target_size=(IM_WIDTH, IM_HEIGHT),
batch_size=batch_size,
class_mode='categorical'
)

# 训练
history_t1 = model.fit_generator(
train_generator,
epochs=1,
steps_per_epoch=10,
validation_data=validation_generator,
validation_steps=10,
class_weight='auto'
)

 model.save('/model/test_model.h5')

# 可用 model.summary() 查看模型结构

#根据模型提取图片特征

target_size = (224,224)

def my_feature(mod, path):
    img = image.load_img(path,target_size=target_size)
    img = image.img_to_array(img)
    img = np.expand_dims(img, axis=0)
    img = preprocess_input(img)
    return mod.predict(img)

 

# 创建模型，获取指定层特征
model_path = './model/my_model.h5'
base_model = load_model(model_path)
model = Model(inputs=base_model.input, outputs=base_model.get_layer('dense_1').output)

 

# 提取特征
img_path = './my_img/bus.jpg'
feat = my_feature(model,img_path) # shape 为 (1,128)
print(feat)
print(feat.shape)

#注意， 当需要提取的图片特征数量较大，比如千万以上，需要的时间是比较长的，这时我们可以采用多核与批处理来进行 (python 由于 GIL 的问题对多线程不友好)。
def pre_processs_image(path):
    if path is not None and os.path.exists(path) and len(path) > 10:
    　　try:
        　　img = cv2.imread(path, cv2.IMREAD_COLOR)
        　　img = cv2.resize(img, (224, 224))
       　　 img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
       　　 img = img.transpose(2, 0, 1)
       　　 return [material_id,img, flag]
   　　 except Exception as err:
        　　traceback.print_exc()
       　　 return None
   　else:
　　　　logging.error('could not find path: ' + path)
　　　　return None

 

#cpu 部分，调用多核处理函数，指定核数为 20
with ProcessPoolExecutor(max_workers=20) as executor:
feat_paras = list(executor.map(pre_processs_image，, material_batch))


# GPU 部分采用批处理
# TODO

 

# create index
d = 128
nlist = 100 # 切分数量
nprobe = 8 # 每次查找分片数量
quantizer_img = faiss.IndexFlatL2(d) #根据欧式距离创建索引

 
image_index = None
model_index = None

if image_feat_array is not None and len(img_feat_list) > 100:
　　image_index = faiss.IndexIVFFlat(quantizer_img, d, nlist, faiss.METRIC_L2)
　　image_index.train(image_feat_array)
　　image_index.add_with_ids(image_feat_array,image_id_array)
　　image_index.nprobe = nprobe
　　image_index.dont_dealloc_me = quantizer_img

# 保存当前索引到指定路径
faiss.write_index(img_index,path)

# 测试当前索引
temp_feat = img_feat_list[1]
res_2 = image_index.search(temp_feat, k=5)
logging.info('image search result is:' + str(res_2))