迁移学习理论与实践
↑ 点击蓝字 关注极市平台
极市导读
随着越来越多的机器学习应用场景的出现,而现有表现比较好的监督学习需要大量的标注数据,标注数据是一项枯燥无味且花费巨大的任务,所以迁移学习受到越来越多的关注。本文阐述了迁移学习的理论知识、基于ResNet的迁移学习实验以及基于resnet50的迁移学习模型。 >>加入极市CV技术交流群,走在计算机视觉的最前沿
迁移学习:深度学习未来五年的驱动力?
迁移学习的使用场景
深度卷积网络的可迁移性
迁移学习的使用方法
基于ResNet的迁移学习实验
# 导入相关模块
import os
import pandas as pd
import numpy as np
import cv2
import matplotlib.pyplot as plt
from PIL import Image
from sklearn.preprocessing import LabelEncoder
from sklearn.model_selection import train_test_split
import keras
from keras.models import Model
from keras.layers import Dense, Activation, Flatten, Dropout
from keras.utils import np_utils
from keras.applications.resnet50 import ResNet50
from tqdm import tqdm
提取数据标签
def generate_csv(path):
labels = pd.DataFrame()
# 目录下每一类别文件夹
items = [f for f in os.listdir(path)]
# 遍历每一类别文件夹
for i in tqdm(items):
# 生成图片完整路径
images = [path + I + '/' + img for img in os.listdir(path+i)]
# 生成两列:图像路径和对应标签
labels_data = pd.DataFrame({'images': images, ‘labels’: i})
# 逐条记录合并
labels = pd.concat((labels, labels_data))
# 打乱顺序
labels = labels.sample(frac=1, random_state=42)
return labels
# 生成标签并查看前5行
labels = generate_csv('./flowers/')
labels.head()
图片预处理
# resize缩放
img = cv2.resize(img, (224, 224))
# 转换成RGB色彩显示
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
plt.imshow(img)
plt.xticks([])
plt.yticks([])
# 定义批量读取并缩放
def read_images(df, resize_dim):
total = 0
images_array = []
# 遍历标签文件中的图像路径
for i in tqdm(df.images):
# 读取并resize
img = cv2.imread(i)
img_resized = cv2.resize(img, resize_dim)
total += 1
# 存入图像数组中
images_array.append(img_resized)
'iamges have resized.')
return images_array
# 批量读取
images_array = read_images(labels, (224, 224))
准备数据
# 转化为图像数组
X = np.array(images_array)
# 标签编码
lbl = LabelEncoder().fit(list(labels['labels'].values))
labels['code_labels']=pd.DataFrame(lbl.transform(list(labels['labels'].values)))
# 分类标签转换
y = np_utils.to_categorical(labels.code_labels.values, 5)
# 划分为训练和验证集
X_train, X_valid, y_train, y_valid =
train_test_split(X,y,test_size=0.2, random_state=42)
# 训练集生成器,中间做一些数据增强
train_datagen = ImageDataGenerator(
rescale=1./255,
rotation_range=40,
width_shift_range=0.4,
height_shift_range=0.4,
shear_range=0.2,
zoom_range=0.3,
horizontal_flip=True
)
# 验证集生成器,无需做数据增强
val_datagen = ImageDataGenerator(
rescale=1./255
)
# 按批次导入训练数据
train_generator = train_datagen.flow(
X_train,
y_train,
batch_size=32
)
# 按批次导入验证数据
val_generator = val_datagen.flow(
X_valid,
y_valid,
batch_size=32
)
基于resnet50的迁移学习模型
# 定义模型构建函数
def flower_model():
base_model=ResNet50(include_top=False,weights='imagenet', input_shape=(224, 224, 3))
# 冻结base_model的层,不参与训练
for layers in base_model.layers:
layers.trainable = False
# base_model的输出并展平
model = Flatten()(base_model.output)
# 添加批归一化层
model = BatchNormalization()(model)
# 全连接层
model=Dense(2048,activation='relu', kernel_initializer=he_normal(seed=42))(model)
# 添加批归一化层
model = BatchNormalization()(model)
# 全连接层
model=Dense(1024,activation='relu', kernel_initializer=he_normal(seed=42))(model)
# 添加批归一化层
model = BatchNormalization()(model)
# 全连接层并指定分类数和softmax激活函数
model = Dense(5, activation='softmax')(model)
model = Model(inputs=base_model.input, outputs=model)
# 指定损失函数、优化器、性能度量标准并编译
model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])
return model
# 调用模型
model = flower_model()
# 使用fit_generator方法执行训练
flower_model.fit_generator(
generator=train_generator,
steps_per_epoch=len(train_data)/32,
epochs=30,
validation_steps=len(val_data)/32,
validation_data=val_generator,
verbose=2
)
推荐阅读
评论