1. 导包

import matplotlib.pyplot as plt

import numpy as np

import pandas as pd

import tensorflow as tf

from sklearn.preprocessing import StandardScaler

from sklearn.model_selection import train_test_split

2. 图像分类 fashion_mnist

数据处理

# 原始数据

(X_train_all, y_train_all),(X_test, y_test) =

tf.keras.datasets.fashion_mnist.load_data()

# 训练集、验证集拆分

X_train, X_valid, y_train, y_valid = train_test_split(X_train_all,

y_train_all, test_size=0.25)

# 数据标准化，你也可以用除以255的方式实现归一化

# 注意最后reshape中的1，代表图像只有一个channel，即当前图像是灰度图

scaler = StandardScaler()

X_train_scaled = scaler.fit_transform(X_train.reshape(-1, 28 *

28)).reshape(-1, 28, 28, 1)

X_valid_scaled = scaler.transform(X_valid.reshape(-1, 28 * 28)).reshape(-1,

28, 28, 1)

X_test_scaled = scaler.transform(X_test.reshape(-1, 28 * 28)).reshape(-1,

28, 28, 1)

构建CNN模型

model = tf.keras.models.Sequential()

# 多个卷积层

model.add(tf.keras.layers.Conv2D(filters=32, kernel_size=[5, 5],

padding="same", activation="relu", input_shape=(28, 28, 1)))

model.add(tf.keras.layers.MaxPool2D(pool_size=[2, 2], strides=2))

model.add(tf.keras.layers.Conv2D(filters=64, kernel_size=[5, 5],

padding="same", activation="relu"))

model.add(tf.keras.layers.MaxPool2D(pool_size=[2, 2], strides=2))

# 将前面卷积层得出的多维数据转为一维

# 7和前面的kernel_size、padding、MaxPool2D有关

# Conv2D: 28*28 -> 28*28 (因为padding="same")

# MaxPool2D: 28*28 -> 14*14

# Conv2D: 14*14 -> 14*14 (因为padding="same")

# MaxPool2D: 14*14 -> 7*7

model.add(tf.keras.layers.Reshape(target_shape=(7 * 7 * 64,)))

# 传入全连接层

model.add(tf.keras.layers.Dense(1024, activation="relu"))

model.add(tf.keras.layers.Dense(10, activation="softmax"))

# compile

model.compile(loss = "sparse_categorical_crossentropy",

optimizer = "sgd",

metrics = ["accuracy"])

模型训练

callbacks = [

tf.keras.callbacks.EarlyStopping(min_delta=1e-3, patience=5)

]

history = model.fit(X_train_scaled, y_train, epochs=15,

validation_data=(X_valid_scaled, y_valid),

callbacks = callbacks)

Train on 50000 samples, validate on 10000 samples

Epoch 1/15

50000/50000 [==============================] - 17s 343us/sample - loss:

0.5707 - accuracy: 0.7965 - val_loss: 0.4631 - val_accuracy: 0.8323

Epoch 2/15

50000/50000 [==============================] - 13s 259us/sample - loss:

0.3728 - accuracy: 0.8669 - val_loss: 0.3573 - val_accuracy: 0.8738

...

Epoch 13/15

50000/50000 [==============================] - 12s 244us/sample - loss:

0.1625 - accuracy: 0.9407 - val_loss: 0.2489 - val_accuracy: 0.9112

Epoch 14/15

50000/50000 [==============================] - 12s 240us/sample - loss:

0.1522 - accuracy: 0.9451 - val_loss: 0.2584 - val_accuracy: 0.9104

Epoch 15/15

50000/50000 [==============================] - 12s 237us/sample - loss:

0.1424 - accuracy: 0.9500 - val_loss: 0.2521 - val_accuracy: 0.9114

作图

def plot_learning_curves(history):

pd.DataFrame(history.history).plot(figsize=(8, 5))

plt.grid(True)

#plt.gca().set_ylim(0, 1)

plt.show()

plot_learning_curves(history)

测试集评估准确率

model.evaluate(X_test_scaled, y_test)

[0.269884311157465, 0.9071]

可以看到使用CNN后，图像分类的准确率明显提升了。之前的模型是0.8747，现在是0.9071。

3. 图像分类 Dogs vs. Cats

3.1 原始数据

原始数据下载

Kaggle: https://www.kaggle.com/c/dogs-vs-cats/

百度网盘: https://pan.baidu.com/s/13hw4LK8ihR6-6-8mpjLKDA 提取码 dmp4

读取一张图片，并展示

image_string =

tf.io.read_file("C:/Users/Skey/Downloads/datasets/cat_vs_dog/train/cat.28.jpg")

image_decoded = tf.image.decode_jpeg(image_string)

plt.imshow(image_decoded)

3.2 利用Dataset加载图片

由于原始图片过多，我们不能将所有图片一次加载入内存。Tensorflow为我们提供了便利的Dataset

API，可以从硬盘中一批一批的加载数据，以用于训练。

处理本地图片路径与标签

# 训练数据的路径

train_dir = "C:/Users/Skey/Downloads/datasets/cat_vs_dog/train/"

train_filenames = [] # 所有图片的文件名

train_labels = [] # 所有图片的标签

for filename in os.listdir(train_dir):

train_filenames.append(train_dir + filename)

if (filename.startswith("cat")):

train_labels.append(0) # 将cat标记为0

else:

train_labels.append(1) # 将dog标记为1

# 数据随机拆分

郑州人流哪家医院做的好 http://www.csyhjlyy.com/

X_train, X_valid, y_train, y_valid = train_test_split(train_filenames,

train_labels, test_size=0.2)

定义一个解码图片的方法

def _decode_and_resize(filename, label):

image_string = tf.io.read_file(filename) # 读取图片

image_decoded = tf.image.decode_jpeg(image_string) # 解码

image_resized = tf.image.resize(image_decoded, [256, 256]) / 255.0 #

重置size，并归一化

return image_resized, label

定义 Dataset，用于加载图片数据

# 训练集

train_dataset = tf.data.Dataset.from_tensor_slices((train_filenames,

train_labels))

train_dataset = train_dataset.map(

map_func=_decode_and_resize, # 调用前面定义的方法，解析filename，转为特征和标签

num_parallel_calls=tf.data.experimental.AUTOTUNE)

train_dataset = train_dataset.shuffle(buffer_size=128) # 设置缓冲区大小

train_dataset = train_dataset.batch(32) # 每批数据的量

train_dataset = train_dataset.prefetch(tf.data.experimental.AUTOTUNE) #

启动预加载图片，也就是说CPU会提前从磁盘加载数据，不用等上一次训练完后再加载

# 验证集

valid_dataset = tf.data.Dataset.from_tensor_slices((valid_filenames,

valid_labels))

valid_dataset = valid_dataset.map(

map_func=_decode_and_resize,

num_parallel_calls=tf.data.experimental.AUTOTUNE)

valid_dataset = valid_dataset.batch(32)

3.3 构建CNN模型，并训练

构建模型与编译

model = tf.keras.Sequential([

# 卷积，32个filter(卷积核)，每个大小为3*3，步长为1

tf.keras.layers.Conv2D(32, 3, activation='relu', input_shape=(256, 256,

3)),

# 池化，默认大小2*2，步长为2

tf.keras.layers.MaxPooling2D(),

tf.keras.layers.Conv2D(32, 5, activation='relu'),

tf.keras.layers.MaxPooling2D(),

tf.keras.layers.Flatten(),

tf.keras.layers.Dense(64, activation='relu'),

tf.keras.layers.Dense(2, activation='softmax')

])

model.compile(

optimizer=tf.keras.optimizers.Adam(learning_rate=0.001),

loss=tf.keras.losses.sparse_categorical_crossentropy,

metrics=[tf.keras.metrics.sparse_categorical_accuracy]

)

模型总览

model.summary()

Model: "sequential_1"

_________________________________________________________________

Layer (type) Output Shape Param #

=================================================================

conv2d_2 (Conv2D) (None, 254, 254, 32) 896

_________________________________________________________________

max_pooling2d_2 (MaxPooling2 (None, 127, 127, 32) 0

_________________________________________________________________

conv2d_3 (Conv2D) (None, 123, 123, 32) 25632

_________________________________________________________________

max_pooling2d_3 (MaxPooling2 (None, 61, 61, 32) 0

_________________________________________________________________

flatten_1 (Flatten) (None, 119072) 0

_________________________________________________________________

dense_2 (Dense) (None, 64) 7620672

_________________________________________________________________

dense_3 (Dense) (None, 2) 130

=================================================================

Total params: 7,647,330

Trainable params: 7,647,330

Non-trainable params: 0

开始训练

model.fit(train_dataset, epochs=10, validation_data=valid_dataset)

由于数据量大，此处训练时间较久

需要注意的是此处打印的step，每个step指的是一个batch(例如32个样本一个batch)

模型评估

test_dataset = tf.data.Dataset.from_tensor_slices((valid_filenames,

valid_labels))

test_dataset = test_dataset.map(_decode_and_resize)

test_dataset = test_dataset.batch(32)

print(model.metrics_names)

print(model.evaluate(test_dataset))