文章目录
卷积运算的原理
输出特征图与输入的宽度和高度不同的原因
- 边界效应，有一部分小框框超出了边界（可以通过对输入进行填充解决）
- 步幅
边界效应解决方法
通过调节padding = ‘same’（默认值为’valid’表示不用填充）对原输入图像进行填充
最大池化运算
-
在每个MaxPooling2D之后，特征图的尺寸都会减半（如从26 * 26 变为13 * 13）这是原于我们会对图片进行下采样
-
最大池化层的下采样通常使用2 * 2的窗口和步幅2， 与此相对的卷积通常使用3 * 3的窗口和步幅1
使用MaxPooling层的原因
- 通过对全局样本进行下采样，可以使得学习模型能够从一个更大的维度去看待整个样本，从局部 --> 全局
- 减少运算开支，防止过拟合
当然，实现下采样的方式有许多种，MaxPooling层的使用只是其中的一种方案，也可以使用平均池化层来代替最大池化，平均池化层志求方块中的平均值而不是最大值，但是最大池化层的效果明显要更好（平均池化层使得区域的特征变得不那么明显，即淡化了图像原有的特征信息）
猫狗识别案例
下面是一个Kaggle上的猫狗识别案例
5-4 数据预处理
import os, shutil
# 读入数据
# 数据原始路径
original_dataset_dir = r'E:\code\PythonDeep\DataSet\dogs-vs-cats\train'
# 我们需要创建的文件根目录
base_dir = r'E:\code\PythonDeep\DataSet\sampledata'
os.mkdir(base_dir)
# 训练集数据文件夹
train_dir = os.path.join(base_dir, "train")
os.mkdir(train_dir)
# 验证集根目录
validation_dir = os.path.join(base_dir, "validation")
os.mkdir(validation_dir)
# 测试集根目录
test_dir = os.path.join(base_dir, "test")
os.mkdir(test_dir)
# 训练集目录下
# 猫图像训练集目录
train_cats_dir = os.path.join(train_dir, 'cats')
os.mkdir(train_cats_dir)
# 狗图像训练集目录
train_dogs_dir = os.path.join(train_dir, 'dogs')
os.mkdir(train_dogs_dir)
# 验证集目录下
# 猫验证集图像目录
validation_cats_dir = os.path.join(validation_dir, 'cats')
os.mkdir(validation_cats_dir)
# 狗验证集图像目录
validation_dogs_dir = os.path.join(validation_dir, 'dogs')
os.mkdir(validation_dogs_dir)
# 测试集目录下
# 猫测试集图像目录
test_cats_dir = os.path.join(test_dir, 'cats')
os.mkdir(test_cats_dir)
# 狗测试集图像目录
test_dogs_dir = os.path.join(test_dir, 'dogs')
os.mkdir(test_dogs_dir)
# =========================================================
# 将前1000张猫的图像复制到train_cats_dir
# 文件名：cat.0.jpg
fnames = ['cat.{}.jpg'.format(i) for i in range(1000)] # 正则表达式
for fname in fnames:
src = os.path.join(original_dataset_dir, fname) # 源地址
dst = os.path.join(train_cats_dir, fname) # 目标地址
shutil.copyfile(src, dst)
# 将500张猫的图片复制到validation_cats_dir
fnames = ['cat.{}.jpg'.format(i) for i in range(1000, 1500)]
for fname in fnames:
src = os.path.join(original_dataset_dir, fname)
dst = os.path.join(validation_cats_dir, fname)
shutil.copyfile(src, dst)
# 将500张猫的图片复制到test_cats_dir
fnames = ['cat.{}.jpg'.format(i) for i in range(1500, 2000)]
for fname in fnames:
src = os.path.join(original_dataset_dir, fname)
dst = os.path.join(test_cats_dir, fname)
shutil.copyfile(src, dst)
# 将1000张狗的图片复制到train_dogs_dir
fnames = ['dog.{}.jpg'.format(i) for i in range(1000)]
for fname in fnames:
src = os.path.join(original_dataset_dir, fname)
dst = os.path.join(train_dogs_dir, fname)
shutil.copyfile(src, dst)
# 将500张狗的图片复制到validation_dogs_dir
fnames = ['dog.{}.jpg'.format(i) for i in range(1000, 1500)]
for fname in fnames:
src = os.path.join(original_dataset_dir, fname)
dst = os.path.join(validation_dogs_dir, fname)
shutil.copyfile(src, dst)
# 将500张狗的图片复制到test_dogs_dir
fnames = ['dog.{}.jpg'.format(i) for i in range(1500, 2000)]
for fname in fnames:
src = os.path.join(original_dataset_dir, fname)
dst = os.path.join(test_dogs_dir, fname)
shutil.copyfile(src, dst)
# 打印，看数据集是否正确
print('total training cat images:', len(os.listdir(train_cats_dir)))
print('total training dog images:', len(os.listdir(train_dogs_dir)))
print('total validation cat images:', len(os.listdir(validation_cats_dir)))
print('total validation dog images:', len(os.listdir(validation_dogs_dir)))
print('total test cat images:', len(os.listdir(test_cats_dir)))
print('total test dog images:', len(os.listdir(test_dogs_dir)))
total training cat images: 1000
total training dog images: 1000
total validation cat images: 500
total validation dog images: 500
total test cat images: 500
total test dog images: 500
构建深度学习网络
- 在卷积深度学习网络中，特征图的深度在逐渐增大，从32到128，而尺寸在逐渐减小，从150 * 150 到 7 * 7
- 由于我们这个问题是一个研究二分类的问题（区别猫和狗），最后的激活层我们使用sigmoid
5-5 将猫狗分类的小心卷积神经网络实例化
# 构建卷积神经网络模型
from keras import layers
from keras import models
model = models.Sequential()
# 卷积层
model.add(layers.Conv2D(32, (3, 3), activation = 'relu', input_shape = (150, 150, 3)))
model.add(layers.MaxPooling2D((2, 2)))
model.add(layers.Conv2D(64, (3, 3), activation = 'relu'))
model.add(layers.MaxPooling2D((2, 2)))
model.add(layers.Conv2D(128, (3, 3), activation = 'relu'))
model.add(layers.MaxPooling2D((2, 2)))
model.add(layers.Conv2D(128, (3, 3), activation = 'relu'))
model.add(layers.MaxPooling2D((2, 2)))
# 全连接层
model.add(layers.Flatten()) # 将输出数据变为一维向量
model.add(layers.Dense(512, activation = 'relu'))
model.add(layers.Dense(1, activation = 'sigmoid'))
# 查看模型概况
model.summary()
Model: "sequential_2"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
conv2d_5 (Conv2D) (None, 148, 148, 32) 896
_________________________________________________________________
max_pooling2d_5 (MaxPooling2 (None, 74, 74, 32) 0
_________________________________________________________________
conv2d_6 (Conv2D) (None, 72, 72, 64) 18496
_________________________________________________________________
max_pooling2d_6 (MaxPooling2 (None, 36, 36, 64) 0
_________________________________________________________________
conv2d_7 (Conv2D) (None, 34, 34, 128) 73856
_________________________________________________________________
max_pooling2d_7 (MaxPooling2 (None, 17, 17, 128) 0
_________________________________________________________________
conv2d_8 (Conv2D) (None, 15, 15, 128) 147584
_________________________________________________________________
max_pooling2d_8 (MaxPooling2 (None, 7, 7, 128) 0
_________________________________________________________________
flatten_2 (Flatten) (None, 6272) 0
_________________________________________________________________
dense_3 (Dense) (None, 512) 3211776
_________________________________________________________________
dense_4 (Dense) (None, 1) 513
=================================================================
Total params: 3,453,121
Trainable params: 3,453,121
Non-trainable params: 0
_________________________________________________________________
5-6 配置模型用于训练
from keras import optimizers
model.compile(loss = 'binary_crossentropy',
optimizer = optimizers.RMSprop(lr = 1e-4),
metrics = ['acc']) # 注意区分l和1，Lr = 一（中文）e减4
5-7 使用ImageDataGenerator从目录读取图像
接下来我们减数据个书画为经过预处理的浮点数张量
- 读取图像文件
- 将JEPG转换为RGB
- 将像素网格转换为浮点数
- 将0到255的数据转换为小数点[0, 1]区间
from keras.preprocessing.image import ImageDataGenerator
train_datagen = ImageDataGenerator(rescale = 1./255) # 让数据每个元素除以255
test_datagen = ImageDataGenerator(rescale = 1./255)
# 调整图像大小
train_generator = train_datagen.flow_from_directory(
train_dir,
target_size = (150, 150),
batch_size = 20,
class_mode = 'binary')
validation_generator = test_datagen.flow_from_directory(
validation_dir,
target_size = (150, 150),
batch_size = 20,
class_mode = 'binary')
# 打印生成器有关信息
for data_batch, labels_batch in train_generator:
print('data batch shape: ', data_batch.shape)
print('labels batch shape: ', labels_batch.shape)
break
Found 2000 images belonging to 2 classes.
Found 1000 images belonging to 2 classes.
data batch shape: (20, 150, 150, 3)
labels batch shape: (20,)
5-8 利用批量生成器拟合模型
# 训练模型
history = model.fit_generator(train_generator, # 一个不断生成数据的训练迭代器
steps_per_epoch = 100, # 每个batch有20个样本，读完2000个要100次
epochs = 30,
validation_data = validation_generator,
validation_steps = 50)
Epoch 1/30
100/100 [==============================] - 42s 418ms/step - loss: 0.6863 - acc: 0.5410 - val_loss: 0.6788 - val_acc: 0.5510
Epoch 2/30
100/100 [==============================] - 39s 389ms/step - loss: 0.6441 - acc: 0.6320 - val_loss: 0.7571 - val_acc: 0.6170
Epoch 3/30
100/100 [==============================] - 39s 385ms/step - loss: 0.5943 - acc: 0.6820 - val_loss: 0.5072 - val_acc: 0.6700
Epoch 4/30
100/100 [==============================] - 39s 394ms/step - loss: 0.5617 - acc: 0.6995 - val_loss: 0.5658 - val_acc: 0.6700
Epoch 5/30
100/100 [==============================] - 39s 392ms/step - loss: 0.5326 - acc: 0.7185 - val_loss: 0.4614 - val_acc: 0.6870
Epoch 6/30
100/100 [==============================] - 39s 394ms/step - loss: 0.5012 - acc: 0.7535 - val_loss: 0.6301 - val_acc: 0.6940
Epoch 7/30
100/100 [==============================] - 40s 397ms/step - loss: 0.4751 - acc: 0.7590 - val_loss: 0.6279 - val_acc: 0.7090
Epoch 8/30
100/100 [==============================] - 40s 398ms/step - loss: 0.4399 - acc: 0.8010 - val_loss: 0.4719 - val_acc: 0.6990
Epoch 9/30
100/100 [==============================] - 42s 419ms/step - loss: 0.4167 - acc: 0.8125 - val_loss: 0.4850 - val_acc: 0.7340
Epoch 10/30
100/100 [==============================] - 43s 426ms/step - loss: 0.3905 - acc: 0.8150 - val_loss: 0.5103 - val_acc: 0.7260
Epoch 11/30
100/100 [==============================] - 40s 397ms/step - loss: 0.3642 - acc: 0.8365 - val_loss: 0.5101 - val_acc: 0.7410
Epoch 12/30
100/100 [==============================] - 42s 416ms/step - loss: 0.3384 - acc: 0.8555 - val_loss: 0.6325 - val_acc: 0.7300
Epoch 13/30
100/100 [==============================] - 42s 416ms/step - loss: 0.3246 - acc: 0.8635 - val_loss: 0.9336 - val_acc: 0.7340
Epoch 14/30
100/100 [==============================] - 39s 389ms/step - loss: 0.2980 - acc: 0.8725 - val_loss: 1.0578 - val_acc: 0.7220
Epoch 15/30
100/100 [==============================] - 39s 390ms/step - loss: 0.2808 - acc: 0.8725 - val_loss: 1.1070 - val_acc: 0.7260
Epoch 16/30
100/100 [==============================] - 39s 390ms/step - loss: 0.2523 - acc: 0.8990 - val_loss: 1.0064 - val_acc: 0.7340
Epoch 17/30
100/100 [==============================] - 40s 400ms/step - loss: 0.2322 - acc: 0.9000 - val_loss: 0.6108 - val_acc: 0.7520
Epoch 18/30
100/100 [==============================] - 40s 396ms/step - loss: 0.2151 - acc: 0.9190 - val_loss: 0.8014 - val_acc: 0.7320
Epoch 19/30
100/100 [==============================] - 39s 391ms/step - loss: 0.1902 - acc: 0.9305 - val_loss: 0.3588 - val_acc: 0.7320
Epoch 20/30
100/100 [==============================] - 39s 391ms/step - loss: 0.1704 - acc: 0.9370 - val_loss: 0.4965 - val_acc: 0.7300
Epoch 21/30
100/100 [==============================] - 39s 388ms/step - loss: 0.1577 - acc: 0.9415 - val_loss: 0.3101 - val_acc: 0.7230
Epoch 22/30
100/100 [==============================] - 39s 392ms/step - loss: 0.1363 - acc: 0.9510 - val_loss: 0.4775 - val_acc: 0.7390
Epoch 23/30
100/100 [==============================] - 39s 389ms/step - loss: 0.1243 - acc: 0.9570 - val_loss: 0.4934 - val_acc: 0.7370
Epoch 24/30
100/100 [==============================] - 41s 413ms/step - loss: 0.1063 - acc: 0.9710 - val_loss: 1.0973 - val_acc: 0.7130
Epoch 25/30
100/100 [==============================] - 40s 396ms/step - loss: 0.0952 - acc: 0.9710 - val_loss: 1.7752 - val_acc: 0.7110
Epoch 26/30
100/100 [==============================] - 39s 390ms/step - loss: 0.0787 - acc: 0.9780 - val_loss: 0.5990 - val_acc: 0.7390
Epoch 27/30
100/100 [==============================] - 41s 411ms/step - loss: 0.0687 - acc: 0.9830 - val_loss: 0.7672 - val_acc: 0.7330
Epoch 28/30
100/100 [==============================] - 44s 437ms/step - loss: 0.0608 - acc: 0.9825 - val_loss: 0.6554 - val_acc: 0.7400
Epoch 29/30
100/100 [==============================] - 41s 407ms/step - loss: 0.0514 - acc: 0.9875 - val_loss: 0.4879 - val_acc: 0.7340
Epoch 30/30
100/100 [==============================] - 40s 399ms/step - loss: 0.0443 - acc: 0.9870 - val_loss: 0.4517 - val_acc: 0.7260
5-9 保存模型
model.save('cats_and_dogs_small_1.h5')
5-10 绘制训练过程中的损失曲线和进度曲线
import matplotlib.pyplot as plt
acc = history.history['acc']
val_acc = history.history['val_acc']
loss = history.history['loss']
val_loss = history.history['val_loss']
epochs = range(1, len(acc) + 1)
# 绘图
plt.plot(epochs, acc, 'bo', label = 'Training acc')
plt.plot(epochs, val_acc, 'b', label = 'Validation acc')
plt.title('Training and validation accuracy')
plt.legend()
plt.figure()
# 绘制图像2
plt.plot(epochs, loss, 'bo', label = 'Training loss')
plt.plot(epochs, val_loss, 'b', label = 'Validation loss')
plt.title('Training and validation loss')
plt.legend()
plt.show()
模型准确度准确曲线
模型损失曲线
总结
从上图来看，我们训练的模型从整体上还是有过拟合的特征，训练精度直接逼近100，而验证精度则一直在70左右，训练精度应该在第5轮左右达到了最小值，为解决过拟合问题，我们在下一节中将使用数据增强的方法。
写在最后
注：本文代码来自《Python 深度学习》，做成电子笔记的方式上传，仅供学习参考，作者均已运行成功，如有遗漏请练习本文作者
各位看官，都看到这里了，麻烦动动手指头给博主来个点赞8，您的支持作者最大的创作动力哟！
＜（＾－＾）＞
才疏学浅，若有纰漏，恳请斧正
本文章仅用于各位同志作为学习交流之用，不作任何商业用途，若涉及版权问题请速与作者联系，望悉知
163
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
