實戰(zhàn) 遷移學(xué)習(xí) VGG19、ResNet50、InceptionV3 實踐貓狗大戰(zhàn) 問題

昵稱56314485 2018-06-03

展開全文

一、實踐流程

1、數(shù)據(jù)預(yù)處理

主要是對訓(xùn)練數(shù)據(jù)進行隨機偏移、轉(zhuǎn)動等變換圖像處理，這樣可以盡可能讓訓(xùn)練數(shù)據(jù)多樣化

另外處理數(shù)據(jù)方式采用分批無序讀取的形式，避免了數(shù)據(jù)按目錄排序訓(xùn)練

[python] view plain copy

print?

#數(shù)據(jù)準備
def DataGen(self, dir_path, img_row, img_col, batch_size, is_train):
if is_train:
datagen = ImageDataGenerator(rescale=1./255,
zoom_range=0.25, rotation_range=15.,
channel_shift_range=25., width_shift_range=0.02, height_shift_range=0.02,
horizontal_flip=True, fill_mode='constant')
else:
datagen = ImageDataGenerator(rescale=1./255)
generator = datagen.flow_from_directory(
dir_path, target_size=(img_row, img_col),
batch_size=batch_size,
shuffle=is_train)
return generator

2、載入現(xiàn)有模型

這個部分是核心工作，目的是使用ImageNet訓(xùn)練出的權(quán)重來做我們的特征提取器，注意這里后面的分類層去掉

[python] view plain copy

print?

base_model = InceptionV3(weights='imagenet', include_top=False, pooling=None,
input_shape=(img_rows, img_cols, color),
classes=nb_classes)

然后是凍結(jié)這些層，因為是訓(xùn)練好的

[python] view plain copy

print?

for layer in base_model.layers:
layer.trainable = False

而分類部分，需要我們根據(jù)現(xiàn)有需求來新定義的，這里可以根據(jù)實際情況自己進行調(diào)整，比如這樣

[python] view plain copy

print?

x = base_model.output
# 添加自己的全鏈接分類層
x = GlobalAveragePooling2D()(x)
x = Dense(1024, activation='relu')(x)
predictions = Dense(nb_classes, activation='softmax')(x)

或者

[python] view plain copy

print?

x = base_model.output
#添加自己的全鏈接分類層
x = Flatten()(x)
predictions = Dense(nb_classes, activation='softmax')(x)

3、訓(xùn)練模型

這里我們用fit_generator函數(shù)，它可以避免了一次性加載大量的數(shù)據(jù)，并且生成器與模型將并行執(zhí)行以提高效率。比如可以在CPU上進行實時的數(shù)據(jù)提升，同時在GPU上進行模型訓(xùn)練

[python] view plain copy

print?

history_ft = model.fit_generator(
train_generator,
steps_per_epoch=steps_per_epoch,
epochs=epochs,
validation_data=validation_generator,
validation_steps=validation_steps)

二、貓狗大戰(zhàn)數(shù)據(jù)集

訓(xùn)練數(shù)據(jù)540M，測試數(shù)據(jù)270M，大家可以去官網(wǎng)下載

https://www./c/dogs-vs-cats-redux-kernels-edition/data

下載后把數(shù)據(jù)分成dog和cat兩個目錄來存放

三、訓(xùn)練

訓(xùn)練的時候會自動去下權(quán)值，比如vgg19_weights_tf_dim_ordering_tf_kernels_notop.h5，但是如果我們已經(jīng)下載好了的話，可以改源代碼，讓他直接讀取我們的下載好的權(quán)值，比如在resnet50.py中

1、VGG19

vgg19的深度有26層，參數(shù)達到了549M，原模型最后有3個全連接層做分類器所以我還是加了一個1024的全連接層，訓(xùn)練10輪的情況達到了89%

2、ResNet50

ResNet50的深度達到了168層，但是參數(shù)只有99M，分類模型我就簡單點，一層直接分類，訓(xùn)練10輪的達到了96%的準確率

3、inception_v3

InceptionV3的深度159層，參數(shù)92M，訓(xùn)練10輪的結(jié)果

這是一層直接分類的結(jié)果

這是加了一個512全連接的，大家可以隨意調(diào)整測試

四、完整的代碼

[python] view plain copy

print?

# -*- coding: utf-8 -*-
import os
from keras.utils import plot_model
from keras.applications.resnet50 import ResNet50
from keras.applications.vgg19 import VGG19
from keras.applications.inception_v3 import InceptionV3
from keras.layers import Dense,Flatten,GlobalAveragePooling2D
from keras.models import Model,load_model
from keras.optimizers import SGD
from keras.preprocessing.image import ImageDataGenerator
import matplotlib.pyplot as plt
class PowerTransferMode:
#數(shù)據(jù)準備
def DataGen(self, dir_path, img_row, img_col, batch_size, is_train):
if is_train:
datagen = ImageDataGenerator(rescale=1./255,
zoom_range=0.25, rotation_range=15.,
channel_shift_range=25., width_shift_range=0.02, height_shift_range=0.02,
horizontal_flip=True, fill_mode='constant')
else:
datagen = ImageDataGenerator(rescale=1./255)
generator = datagen.flow_from_directory(
dir_path, target_size=(img_row, img_col),
batch_size=batch_size,
#class_mode='binary',
shuffle=is_train)
return generator
#ResNet模型
def ResNet50_model(self, lr=0.005, decay=1e-6, momentum=0.9, nb_classes=2, img_rows=197, img_cols=197, RGB=True, is_plot_model=False):
color = 3 if RGB else 1
base_model = ResNet50(weights='imagenet', include_top=False, pooling=None, input_shape=(img_rows, img_cols, color),
classes=nb_classes)
#凍結(jié)base_model所有層，這樣就可以正確獲得bottleneck特征
for layer in base_model.layers:
layer.trainable = False
x = base_model.output
#添加自己的全鏈接分類層
x = Flatten()(x)
#x = GlobalAveragePooling2D()(x)
#x = Dense(1024, activation='relu')(x)
predictions = Dense(nb_classes, activation='softmax')(x)
#訓(xùn)練模型
model = Model(inputs=base_model.input, outputs=predictions)
sgd = SGD(lr=lr, decay=decay, momentum=momentum, nesterov=True)
model.compile(loss='categorical_crossentropy', optimizer=sgd, metrics=['accuracy'])
#繪制模型
if is_plot_model:
plot_model(model, to_file='resnet50_model.png',show_shapes=True)
return model
#VGG模型
def VGG19_model(self, lr=0.005, decay=1e-6, momentum=0.9, nb_classes=2, img_rows=197, img_cols=197, RGB=True, is_plot_model=False):
color = 3 if RGB else 1
base_model = VGG19(weights='imagenet', include_top=False, pooling=None, input_shape=(img_rows, img_cols, color),
classes=nb_classes)
#凍結(jié)base_model所有層，這樣就可以正確獲得bottleneck特征
for layer in base_model.layers:
layer.trainable = False
x = base_model.output
#添加自己的全鏈接分類層
x = GlobalAveragePooling2D()(x)
x = Dense(1024, activation='relu')(x)
predictions = Dense(nb_classes, activation='softmax')(x)
#訓(xùn)練模型
model = Model(inputs=base_model.input, outputs=predictions)
sgd = SGD(lr=lr, decay=decay, momentum=momentum, nesterov=True)
model.compile(loss='categorical_crossentropy', optimizer=sgd, metrics=['accuracy'])
# 繪圖
if is_plot_model:
plot_model(model, to_file='vgg19_model.png',show_shapes=True)
return model
# InceptionV3模型
def InceptionV3_model(self, lr=0.005, decay=1e-6, momentum=0.9, nb_classes=2, img_rows=197, img_cols=197, RGB=True,
is_plot_model=False):
color = 3 if RGB else 1
base_model = InceptionV3(weights='imagenet', include_top=False, pooling=None,
input_shape=(img_rows, img_cols, color),
classes=nb_classes)
# 凍結(jié)base_model所有層，這樣就可以正確獲得bottleneck特征
for layer in base_model.layers:
layer.trainable = False
x = base_model.output
# 添加自己的全鏈接分類層
x = GlobalAveragePooling2D()(x)
x = Dense(1024, activation='relu')(x)
predictions = Dense(nb_classes, activation='softmax')(x)
# 訓(xùn)練模型
model = Model(inputs=base_model.input, outputs=predictions)
sgd = SGD(lr=lr, decay=decay, momentum=momentum, nesterov=True)
model.compile(loss='categorical_crossentropy', optimizer=sgd, metrics=['accuracy'])
# 繪圖
if is_plot_model:
plot_model(model, to_file='inception_v3_model.png', show_shapes=True)
return model
#訓(xùn)練模型
def train_model(self, model, epochs, train_generator, steps_per_epoch, validation_generator, validation_steps, model_url, is_load_model=False):
# 載入模型
if is_load_model and os.path.exists(model_url):
model = load_model(model_url)
history_ft = model.fit_generator(
train_generator,
steps_per_epoch=steps_per_epoch,
epochs=epochs,
validation_data=validation_generator,
validation_steps=validation_steps)
# 模型保存
model.save(model_url,overwrite=True)
return history_ft
# 畫圖
def plot_training(self, history):
acc = history.history['acc']
val_acc = history.history['val_acc']
loss = history.history['loss']
val_loss = history.history['val_loss']
epochs = range(len(acc))
plt.plot(epochs, acc, 'b-')
plt.plot(epochs, val_acc, 'r')
plt.title('Training and validation accuracy')
plt.figure()
plt.plot(epochs, loss, 'b-')
plt.plot(epochs, val_loss, 'r-')
plt.title('Training and validation loss')
plt.show()
if __name__ == '__main__':
image_size = 197
batch_size = 32
transfer = PowerTransferMode()
#得到數(shù)據(jù)
train_generator = transfer.DataGen('data/cat_dog_Dataset/train', image_size, image_size, batch_size, True)
validation_generator = transfer.DataGen('data/cat_dog_Dataset/test', image_size, image_size, batch_size, False)
#VGG19
#model = transfer.VGG19_model(nb_classes=2, img_rows=image_size, img_cols=image_size, is_plot_model=False)
#history_ft = transfer.train_model(model, 10, train_generator, 600, validation_generator, 60, 'vgg19_model_weights.h5', is_load_model=False)
#ResNet50
model = transfer.ResNet50_model(nb_classes=2, img_rows=image_size, img_cols=image_size, is_plot_model=False)
history_ft = transfer.train_model(model, 10, train_generator, 600, validation_generator, 60, 'resnet50_model_weights.h5', is_load_model=False)
#InceptionV3
#model = transfer.InceptionV3_model(nb_classes=2, img_rows=image_size, img_cols=image_size, is_plot_model=True)
#history_ft = transfer.train_model(model, 10, train_generator, 600, validation_generator, 60, 'inception_v3_model_weights.h5', is_load_model=False)
# 訓(xùn)練的acc_loss圖
transfer.plot_training(history_ft)