/
- 案例效果
2. 数据集以及迁移需求
数据集是某场景下5个类别图片的识别
我们利用现有的VGG模型去进行微调
- 思路和步骤
- 读取本地的图片数据以及类别
- keras.preprocessing.image import ImageDataGenerator提供了读取转换功能
- 模型的结构修改（添加我们自定的分类层）
- freeze掉原始VGG模型
- 编译以及训练和保存模型方式
- 输入数据进行预测
train_datagen = ImageDataGenerator(
rescale=1./255,
shear_range=0.2,
zoom_range=0.2,
horizontal_flip=True)
test_datagen = ImageDataGenerator(rescale=1./255)
train_generator = train_datagen.flow_from_directory(
'data/train',
target_size=(150, 150),
batch_size=32,
class_mode='binary')
validation_generator = test_datagen.flow_from_directory(
'data/validation',
target_size=(150, 150),
batch_size=32,
class_mode='binary')
使用fit_generator
model.fit_generator(
train_generator,
steps_per_epoch=2000,
epochs=50,
validation_data=validation_generator,
validation_steps=800)
def __init__(self):
# 定义训练和测试图片的变化方式，标准化以及数据增强
self.train_generator = ImageDataGenerator(rescale=1.0 / 255)
self.test_generator = ImageDataGenerator(rescale=1.0 / 255)
# 指定训练集和测试集的目录
self.train_dir = "./data/train"
self.test_dir = "./data/test"
# 定义图片训练的相关网络参数
self.image_size = (224, 224)
self.batch_size = 32
def get_local_data(self):
"""
读取本地的图片数据以及类别
:return:训练数据和测试数据迭代器
"""
# 使用flow_from_derectory
train_gen = self.train_generator.flow_from_directory(self.train_dir,
target_size=self.image_size,
batch_size=self.batch_size,
class_mode='binary',
shuffle=True)
test_gen = self.test_generator.flow_from_directory(self.test_dir,
target_size=self.image_size,
batch_size=self.batch_size,
class_mode='binary',
shuffle=True)
return train_gen, test_gen
# 定义迁移学习的基类模型 # 不包含VGG当中3个全连接层的模型加载并且加载了参数 self.batch_model = VGG16(weights='imagenet', include_top=False)
模型源码：
if include_top:
# Classification block
x = layers.Flatten(name='flatten')(x)
x = layers.Dense(4096, activation='relu', name='fc1')(x)
x = layers.Dense(4096, activation='relu', name='fc2')(x)
x = layers.Dense(classes, activation='softmax', name='predictions')(x)
else:
if pooling == 'avg':
x = layers.GlobalAveragePooling2D()(x)
elif pooling == 'max':
x = layers.GlobalMaxPooling2D()(x)
- 使用前提：迁移学习不需要大量的参数去微调
def refine_vgg_model(self):
添加尾部全连接层
:return:
# [<tf.Tensor 'block5_pool/MaxPool:0' shape=(?, ?, ?, 512) dtype=float32>]
x = self.base_model.outputs[0]
# 输出到全连接层，加上全局池化 [None, ?, ?, 512]---->[None, 1 * 512]
x = keras.layers.GlobalAveragePooling2D()(x)
x = keras.layers.Dense(1024, activation=tf.nn.relu)(x)
y_predict = keras.layers.Dense(5, activation=tf.nn.softmax)(x)
model = keras.Model(inputs=self.base_model.inputs, outputs=y_predict)
return model
keras.callbacks.ModelCheckpoint(filepath, monitor='val_loss', verbose=0, save_best_only=False, save_weights_only=False, mode='auto', period=1)
modelckpt = keras.callbacks.ModelCheckpoint('./ckpt/fine_tuning/weights.{epoch:02d}.hdf5',
monitor='val_acc',
save_weights_only=True,
save_best_only=True,
mode='auto',
period=1)
model.fit_generator(train_gen, epochs=3, validation_data=test_gen, callbacks=[modelckpt])
def fit_generator(self, model, train_gen, test_gen):
"""
训练模型，model.fit_generator()不是选择model.fit()
:param model:
:param train_gen:
:param test_gen:
:return:
"""
# 每一次迭代准确率记录的h5文件
modelckpt = keras.callbacks.ModelCheckpoint('./hdf5/transfer_{epoch:02d}-{accuracy:.2f}.h5',
monitor='accuracy',
save_weights_only=True,
save_best_only=True,
mode='auto',
period=1)
model.fit_generator(train_gen, epochs=3, validation_data=test_gen, callbacks=[modelckpt])
# 保存模型
# model.save_weights("./model/Transfer.h5")
return Nonefit_generator()可以使用callbacks中的ModelCheckpoints
if __name__ == "__main__":
tm = TransferModel()
train_gen, test_gen = tm.get_local_data()
model = tm.refine_base_model()
tm.freeeze_model()
tm.compile(model)
tm.fit_generator(model, train_gen, test_gen)
def predict(self, model):
'''
预测类型
:param model:
:return:
'''
# 加载模型，transfer_model
model.load_weights("./hdf5/transfer_03-0.98.h5")
# 读取图片， 处理
image = load_img("./data/test/dinosaurs/400.jpg", target_size=(224, 224))
image = img_to_array(image)
# 三维转换为4维 （224, 224, 3）--->(1, 224, 224, 3)
img = image.reshape([1, image.shape[0], image.shape[1], image.shape[2]])
print(img)
# 预测结果进行处理
image = preprocess_input(img)
predictions = model.predict(image)
print(predictions)
res = np.argmax(predictions, axis=1)
print(self.label_dict[str(res[0])])
完整代码：
import numpy as np
import tensorflow as tf
from tensorflow import keras
from tensorflow.keras.preprocessing.image import ImageDataGenerator, load_img, img_to_array
from tensorflow.keras.applications.vgg16 import VGG16, preprocess_input
tf.compat.v1.disable_eager_execution()
class TransferModel(object):
def __init__(self):
# 定义训练和测试图片的变化方式，标准化以及数据增强
self.train_generator = ImageDataGenerator(rescale=1.0 / 255)
self.test_generator = ImageDataGenerator(rescale=1.0 / 255)
# 指定训练集和测试集的目录
self.train_dir = "./data/train"
self.test_dir = "./data/test"
# 定义图片训练的相关网络参数
self.image_size = (224, 224)
self.batch_size = 32
# 定义迁移学习的基类模型
# 不包含VGG当中3个全连接层的模型加载并且加载了参数
self.base_model = VGG16(weights='imagenet', include_top=False)
self.label_dict = {
'0': 'bus',
'1': 'dinosaurs',
'2': 'eleplants',
'3': 'flowers',
'4': 'horse'
}
def get_local_data(self):
"""
读取本地的图片数据以及类别
:return:训练数据和测试数据迭代器
"""
# 使用flow_from_derectory
train_gen = self.train_generator.flow_from_directory(self.train_dir,
target_size=self.image_size,
batch_size=self.batch_size,
class_mode='binary',
shuffle=True)
test_gen = self.test_generator.flow_from_directory(self.test_dir,
target_size=self.image_size,
batch_size=self.batch_size,
class_mode='binary',
shuffle=True)
return train_gen, test_gen
def refine_base_model(self):
"""
微调VGG结构，5blocks后面+全局平均池化（减少迁移学习的参数数量）+两个全连接层
:return:
"""
# 1. 获取原notop模型得出
# [?, ?, ？， 512]
x = self.base_model.outputs[0]
# 2。 在输出后面增加我们的结构
# [?, ?, ？， 512]----》[?, 1*1*512]
x = keras.layers.GlobalAveragePooling2D()(x)
# 3. 定义新的迁移模型
x = keras.layers.Dense(1024, activation=tf.compat.v1.nn.relu)(x)
y_predict = keras.layers.Dense(5, activation=tf.compat.v1.nn.softmax)(x)
# model定义新模型
# VGG模型的输入，输出，y_predict
transfer_model = keras.models.Model(inputs=self.base_model.inputs, outputs=y_predict)
return transfer_model
def freeeze_model(self):
"""
冻结VGG模型（5blocks）
冻结VGG的多少根据数据量
:return:
"""
# 获取所有层， 返回层的列表
for layer in self.base_model.layers:
layer.trainable = False
def compile(self, model):
'''
编译模型
:return:
'''
model.compile(optimizer=keras.optimizers.Adam(),
loss=keras.losses.sparse_categorical_crossentropy,
metrics=['accuracy'])
return None
def fit_generator(self, model, train_gen, test_gen):
"""
训练模型，model.fit_generator()不是选择model.fit()
:param model:
:param train_gen:
:param test_gen:
:return:
"""
# 每一次迭代准确率记录的h5文件
modelckpt = keras.callbacks.ModelCheckpoint('./hdf5/transfer_{epoch:02d}-{accuracy:.2f}.h5',
monitor='accuracy',
save_weights_only=True,
save_best_only=True,
mode='auto',
period=1)
model.fit_generator(train_gen, epochs=3, validation_data=test_gen, callbacks=[modelckpt])
# 保存模型
# model.save_weights("./model/Transfer.h5")
return None
def predict(self, model):
'''
预测类型
:param model:
:return:
'''
# 加载模型，transfer_model
model.load_weights("./hdf5/transfer_03-0.98.h5")
# 读取图片， 处理
image = load_img("./data/test/dinosaurs/400.jpg", target_size=(224, 224))
image = img_to_array(image)
# 三维转换为4维 （224, 224, 3）--->(1, 224, 224, 3)
img = image.reshape([1, image.shape[0], image.shape[1], image.shape[2]])
print(img)
# 预测结果进行处理
image = preprocess_input(img)
predictions = model.predict(image)
print(predictions)
res = np.argmax(predictions, axis=1)
print(self.label_dict[str(res[0])])
if __name__ == "__main__":
tm = TransferModel()
# 训练
# train_gen, test_gen = tm.get_local_data()
# # print(train_gen, test_gen)
# # # for data in train_gen:
# # # print(train_gen) print(tm.batch_model.summary())
# model = tm.refine_base_model()
# # print(model)
# tm.freeeze_model()
# tm.compile(model)
# tm.fit_generator(model, train_gen, test_gen)
"""测试"""
model = tm.refine_base_model()
tm.predict(model)
1888
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
