29- 迁移学习 (TensorFlow系列) (深度学习)

知识要点

• 迁移学习: 使用别人预训练模型参数时，要注意别人的预处理方式。

• 常见的迁移学习方式：

  • 载入权重后训练所有参数.
  • 载入权重后只训练最后几层参数.
  • 载入权重后在原网络基础上再添加一层全连接层，仅训练最后一个全连接层.
• 训练数据是 10_monkeys 数据: 10种猴子的图片集
• 图片显示: plt.imshow(mokey)
• 读取图片: mokey = plt.imread('./50.jpg')

导入resnet 模型:

• resnet50 = keras.applications.ResNet50(include_top=False, pooling='avg')      # 导入模型
• model = keras.models.Sequential()      # 开始建模
• model.add(resnet50)     # 添加resnet 网络
• model.add(keras.layers.Dense(num_classes=10, activation = 'softmax'))     # 添加全连接层
• model.layers[0].trainable = False     # 除了最后一个全连接层, 其余部分参数不变
• 模型配置:

model.compile(loss = 'categorical_crossentropy',
              optimizer = 'adam',
              metrics = ['acc'])

• valid_datagen = keras.preprocessing.image.ImageDataGenerator(preprocessing_function = keras.applications.resnet50.preprocess_input)    # 数据初始化处理
• 指定后面几层参数变化:

# 切片指定, 不可调整的层数
for layer in resnet50.layers[0:-5]:
    layer.trainable = False

---

一 迁移学习

1.1 简介

使用迁移学习的优势：

• 能够快速的训练出一个理想的结果
• 当数据集较小时也能训练出理想的效果

注意：使用别人预训练模型参数时，要注意别人的预处理方式。

1.2 常见迁移方式

常见的迁移学习方式：

• 载入权重后训练所有参数.
• 载入权重后只训练最后几层参数.
• 载入权重后在原网络基础上再添加一层全连接层，仅训练最后一个全连接层.

二 代码实现

2.1 导包

from tensorflow import keras
import numpy as np
import pandas as pd
import tensorflow as tf
import matplotlib.pyplot as plt

cpu=tf.config.list_physical_devices("CPU")
tf.config.set_visible_devices(cpu)
print(tf.config.list_logical_devices())

2.2 迁移模型  (在迁移模型 后加一层)

resnet50 = keras.applications.ResNet50(include_top=False, pooling='avg')

num_classes =10
model = keras.models.Sequential()
model.add(resnet50)
model.add(keras.layers.Dense(num_classes, activation = 'softmax'))
model.summary()

2.3 配置模型 (除最后一层外, 其余参数全部冻结)

# 把除最后一层的参数外, 全部冻结
model.layers[0].trainable = False
model.compile(loss = 'categorical_crossentropy',
              optimizer = 'adam',
              metrics = ['acc'])

2.4 导入数据

train_dir = '../day 48 resnet/training/training/'
valid_dir = '../day 48 resnet/validation/validation/'

• 原始数据处理

train_datagen = keras.preprocessing.image.ImageDataGenerator(
    preprocessing_function = keras.applications.resnet50.preprocess_input,
    rotation_range = 40,
    width_shift_range = 0.2,
    height_shift_range = 0.2,
    shear_range = 0.2,
    zoom_range = 0.2,
    horizontal_flip = True,
    vertical_flip = True,
    fill_mode = 'nearest')

height = 224
width = 224
channels = 3
batch_size = 32
num_classes = 10

train_generator = train_datagen.flow_from_directory(train_dir,
                                                    target_size= (height, width),
                                                    batch_size = batch_size,
                                                    shuffle= True,
                                                    seed = 7,
                                                    class_mode= 'categorical')

valid_datagen = keras.preprocessing.image.ImageDataGenerator(
    preprocessing_function = keras.applications.resnet50.preprocess_input)

valid_generator = valid_datagen.flow_from_directory(valid_dir,
                                                    target_size= (height, width),
                                                    batch_size= batch_size,
                                                    shuffle= True,
                                                    seed = 7,
                                                    class_mode= 'categorical')
print(train_generator.samples)   # 1098
print(valid_generator.samples)   # 272

2.5 模型训练

# 使用迁移学习, 效果较差, 原始数据的处理方式不同
# 修改需处理方式继续执行, 效果较好
histroy = model.fit(train_generator,
                          steps_per_epoch= train_generator.samples // batch_size,
                          epochs = 10,
                          validation_data = valid_generator,
                          validation_steps= valid_generator.samples // batch_size)

 2.6 训练后面几层神经网络参数

resnet50 = keras.applications.ResNet50(include_top=False, pooling='avg', weights='imagenet')

# 切片指定, 不可调整的层数
for layer in resnet50.layers[0:-5]:
    layer.trainable = False
    
# 添加输出层
resnet50_new = keras.models.Sequential([resnet50, keras.layers.Dense(10, activation = 'softmax')])
resnet50_new.compile(loss = 'categorical_crossentropy',
                     optimizer = 'adam',
                     metrics = ['acc'])
resnet50_new.summary()

histroy = resnet50_new.fit(train_generator,
                          steps_per_epoch= train_generator.samples // batch_size,
                          epochs = 10,
                          validation_data = valid_generator,
                          validation_steps= valid_generator.samples // batch_size)

三 图片处理查看

3.1 图片显示

# 预测数据
mokey = plt.imread('./n5020.jpg')
plt.imshow(mokey)    # mokey.shape  (600, 336, 3)

for i in range(2):
    x, y = train_generator.next()
    print(type(x), type(y))    # <class 'numpy.ndarray'> <class 'numpy.ndarray'>
    print('***', x.shape, y.shape)  # *** (32, 224, 224, 3) (32, 10)

3.2 尺寸变换

# 主要是形状和尺寸不对
# 改变尺寸, 再改变形状reshape
from scipy import ndimage  # 专门处理图片
# 改变形状
# 224 = 367 * x  x = 224/367
# 224 = 550 * y  y = 224/550
zoom = (224/mokey.shape[0], 224/mokey.shape[1])
monkey_zoomed = ndimage.zoom(mokey, (224/mokey.shape[0], 224/mokey.shape[1], 1))
monkey_zoomed.shape   # (224, 224, 3)

monkey_1 = keras.applications.resnet50.preprocess_input(monkey_zoomed)
monkey_1.min()    # -123.68

monkey_1 = monkey_1.reshape(1, 224, 224, 3)
model.predict(monkey_1).argmax(axis = 1)  # array([5], dtype=int64)

3.3 resnet 图片处理方式

3.3.1 前景查看

mokey1 = mokey/127.5
plt.imshow(mokey1)

3.3.2 背景查看

mokey1 = mokey1 - 1
plt.imshow(mokey1)

mokey1