人工智能 – Python技术交流与分享

使用transformer进行图像分类

使用Transformer来提升模型的性能
最近几年，Transformer体系结构已成为自然语言处理任务的实际标准，
但其在计算机视觉中的应用还受到限制。在视觉上，注意力要么与卷积网络结合使用，
要么用于替换卷积网络的某些组件，同时将其整体结构保持在适当的位置。2020年10月22日，谷歌人工智能研究院发表一篇题为“An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale”的文章。文章将图像切割成一个个图像块,组成序列化的数据输入Transformer执行图像分类任务。当对大量数据进行预训练并将其传输到多个中型或小型图像识别数据集（如ImageNet、CIFAR-100、VTAB等）时，与目前的卷积网络相比，Vision Transformer（ViT）获得了出色的结果，同时所需的计算资源也大大减少。
这里我们以ViT我模型，实现对数据CiFar10的分类工作，模型性能得到进一步的提升。

1、导入模型

import os
import math
import numpy as np
import pickle as p
import tensorflow as tf
from tensorflow import keras
import matplotlib.pyplot as plt
from tensorflow.keras import layers
import tensorflow_addons as tfa
%matplotlib inline

import os

import math

import numpy as np

import pickle as p

import tensorflow as tf

from tensorflow import keras

import matplotlib.pyplot as plt

from tensorflow.keras import layers

import tensorflow_addons as tfa

%matplotlib inline

这里使用了TensorFlow_addons模块，它实现了核心 TensorFlow 中未提供的新功能。
tensorflow_addons的安装要注意与tf的版本对应关系，请参考：
https://github.com/tensorflow/addons。
安装addons时要注意其版本与tensorflow版本的对应，具体关系以上这个链接有。

2、定义加载函数

def load_CIFAR_data(data_dir):
    """load CIFAR data"""
 
    images_train=[]
    labels_train=[]
    for i in range(5):
        f=os.path.join(data_dir,'data_batch_%d' % (i+1))
        print('loading ',f)
        # 调用 load_CIFAR_batch( )获得批量的图像及其对应的标签
        image_batch,label_batch=load_CIFAR_batch(f)
        images_train.append(image_batch)
        labels_train.append(label_batch)
        Xtrain=np.concatenate(images_train)
        Ytrain=np.concatenate(labels_train)
        del image_batch ,label_batch
    
    Xtest,Ytest=load_CIFAR_batch(os.path.join(data_dir,'test_batch'))
    print('finished loadding CIFAR-10 data')
    
    # 返回训练集的图像和标签，测试集的图像和标签
return (Xtrain,Ytrain),(Xtest,Ytest)

def load_CIFAR_data(data_dir):

"""load CIFAR data"""

images_train=[]

labels_train=[]

for i in range(5):

f=os.path.join(data_dir,'data_batch_%d' % (i+1))

print('loading ',f)

# 调用 load_CIFAR_batch( )获得批量的图像及其对应的标签

image_batch,label_batch=load_CIFAR_batch(f)

images_train.append(image_batch)

labels_train.append(label_batch)

Xtrain=np.concatenate(images_train)

Ytrain=np.concatenate(labels_train)

del image_batch ,label_batch

Xtest,Ytest=load_CIFAR_batch(os.path.join(data_dir,'test_batch'))

print('finished loadding CIFAR-10 data')

# 返回训练集的图像和标签，测试集的图像和标签

return (Xtrain,Ytrain),(Xtest,Ytest)

3、定义批量加载函数

def load_CIFAR_batch(filename):
    """ load single batch of cifar """  
    with open(filename, 'rb')as f:
        # 一个样本由标签和图像数据组成
        #  (3072=32x32x3)
        # ...
        # 
        data_dict = p.load(f, encoding='bytes')
        images= data_dict[b'data']
        labels = data_dict[b'labels']
                
        # 把原始数据结构调整为: BCWH
        images = images.reshape(10000, 3, 32, 32)
        # tensorflow处理图像数据的结构：BWHC
        # 把通道数据C移动到最后一个维度
        images = images.transpose (0,2,3,1)
     
        labels = np.array(labels)
        
        return images, labels

def load_CIFAR_batch(filename):

""" load single batch of cifar """

with open(filename, 'rb')as f:

# 一个样本由标签和图像数据组成

# (3072=32x32x3)

# ...

data_dict = p.load(f, encoding='bytes')

images= data_dict[b'data']

labels = data_dict[b'labels']

# 把原始数据结构调整为: BCWH

images = images.reshape(10000, 3, 32, 32)

# tensorflow处理图像数据的结构：BWHC

# 把通道数据C移动到最后一个维度

images = images.transpose (0,2,3,1)

labels = np.array(labels)

return images, labels

4、加载数据

data_dir = r'C:\Users\wumg\jupyter-ipynb\data\cifar-10-batches-py'
(x_train,y_train),(x_test,y_test) = load_CIFAR_data(data_dir)

1 2	data_dir = r'C:\Users\wumg\jupyter-ipynb\data\cifar-10-batches-py' (x_train,y_train),(x_test,y_test) = load_CIFAR_data(data_dir)

把数据转换为dataset格式

train_dataset = tf.data.Dataset.from_tensor_slices((x_train, y_train))
test_dataset = tf.data.Dataset.from_tensor_slices((x_test, y_test))

1 2	train_dataset = tf.data.Dataset.from_tensor_slices((x_train, y_train)) test_dataset = tf.data.Dataset.from_tensor_slices((x_test, y_test))

5、定义数据预处理及训练模型的一些超参数

num_classes = 10
input_shape = (32, 32, 3)


learning_rate = 0.001
weight_decay = 0.0001
batch_size = 256
num_epochs = 10
image_size = 72  # We'll resize input images to this size
patch_size = 6  # Size of the patches to be extract from the input images
num_patches = (image_size // patch_size) ** 2
projection_dim = 64
num_heads = 4
transformer_units = [
    projection_dim * 2,
    projection_dim,
]  # Size of the transformer layers
transformer_layers = 8
mlp_head_units = [2048, 1024]  # Size of the dense layers of the final classifier

num_classes = 10

input_shape = (32, 32, 3)

learning_rate = 0.001

weight_decay = 0.0001

batch_size = 256

num_epochs = 10

image_size = 72 # We'll resize input images to this size

patch_size = 6 # Size of the patches to be extract from the input images

num_patches = (image_size // patch_size) ** 2

projection_dim = 64

num_heads = 4

transformer_units = [

projection_dim * 2,

projection_dim,

] # Size of the transformer layers

transformer_layers = 8

mlp_head_units = [2048, 1024] # Size of the dense layers of the final classifier

6、定义数据增强模型

data_augmentation = keras.Sequential(
    [
        layers.experimental.preprocessing.Normalization(),
        layers.experimental.preprocessing.Resizing(image_size, image_size),
        layers.experimental.preprocessing.RandomFlip("horizontal"),
        layers.experimental.preprocessing.RandomRotation(factor=0.02),
        layers.experimental.preprocessing.RandomZoom(
            height_factor=0.2, width_factor=0.2
        ),
    ],
    name="data_augmentation",
)
# 使预处理层的状态与正在传递的数据相匹配
#Compute the mean and the variance of the training data for normalization.
data_augmentation.layers[0].adapt(x_train)

data_augmentation = keras.Sequential(

[

layers.experimental.preprocessing.Normalization(),

layers.experimental.preprocessing.Resizing(image_size, image_size),

layers.experimental.preprocessing.RandomFlip("horizontal"),

layers.experimental.preprocessing.RandomRotation(factor=0.02),

layers.experimental.preprocessing.RandomZoom(

height_factor=0.2, width_factor=0.2

name="data_augmentation",

)

# 使预处理层的状态与正在传递的数据相匹配

#Compute the mean and the variance of the training data for normalization.

data_augmentation.layers[0].adapt(x_train)

预处理层是在模型训练开始之前计算其状态的层。他们在训练期间不会得到更新。大多数预处理层为状态计算实现了adapt（）方法。
adapt(data, batch_size=None, steps=None, reset_state=True)该函数参数说明如下：

7、构建模型

7.1 构建多层感知器（MLP）

def mlp(x, hidden_units, dropout_rate):
    for units in hidden_units:
        x = layers.Dense(units, activation=tf.nn.gelu)(x)
        x = layers.Dropout(dropout_rate)(x)
    return x

def mlp(x, hidden_units, dropout_rate):

for units in hidden_units:

x = layers.Dense(units, activation=tf.nn.gelu)(x)

x = layers.Dropout(dropout_rate)(x)

return x

7.2 创建一个类似卷积层的patch层

class Patches(layers.Layer):
    def __init__(self, patch_size):
        super(Patches, self).__init__()
        self.patch_size = patch_size

    def call(self, images):
        batch_size = tf.shape(images)[0]
        patches = tf.image.extract_patches(
            images=images,
            sizes=[1, self.patch_size, self.patch_size, 1],
            strides=[1, self.patch_size, self.patch_size, 1],
            rates=[1, 1, 1, 1],
            padding="VALID",
        )
        patch_dims = patches.shape[-1]
        patches = tf.reshape(patches, [batch_size, -1, patch_dims])
        return patches

class Patches(layers.Layer):

def __init__(self, patch_size):

super(Patches, self).__init__()

self.patch_size = patch_size

def call(self, images):

batch_size = tf.shape(images)[0]

patches = tf.image.extract_patches(

images=images,

sizes=[1, self.patch_size, self.patch_size, 1],

strides=[1, self.patch_size, self.patch_size, 1],

rates=[1, 1, 1, 1],

padding="VALID",

)

patch_dims = patches.shape[-1]

patches = tf.reshape(patches, [batch_size, -1, patch_dims])

return patches

7.3 查看由patch层随机生成的图像块

import matplotlib.pyplot as plt

plt.figure(figsize=(4, 4))
image = x_train[np.random.choice(range(x_train.shape[0]))]
plt.imshow(image.astype("uint8"))
plt.axis("off")

resized_image = tf.image.resize(
    tf.convert_to_tensor([image]), size=(image_size, image_size)
)
patches = Patches(patch_size)(resized_image)
print(f"Image size: {image_size} X {image_size}")
print(f"Patch size: {patch_size} X {patch_size}")
print(f"Patches per image: {patches.shape[1]}")
print(f"Elements per patch: {patches.shape[-1]}")

n = int(np.sqrt(patches.shape[1]))
plt.figure(figsize=(4, 4))
for i, patch in enumerate(patches[0]):
    ax = plt.subplot(n, n, i + 1)
    patch_img = tf.reshape(patch, (patch_size, patch_size, 3))
    plt.imshow(patch_img.numpy().astype("uint8"))
    plt.axis("off")

import matplotlib.pyplot as plt

plt.figure(figsize=(4, 4))

image = x_train[np.random.choice(range(x_train.shape[0]))]

plt.imshow(image.astype("uint8"))

plt.axis("off")

resized_image = tf.image.resize(

tf.convert_to_tensor([image]), size=(image_size, image_size)

)

patches = Patches(patch_size)(resized_image)

print(f"Image size: {image_size} X {image_size}")

print(f"Patch size: {patch_size} X {patch_size}")

print(f"Patches per image: {patches.shape[1]}")

print(f"Elements per patch: {patches.shape[-1]}")

n = int(np.sqrt(patches.shape[1]))

plt.figure(figsize=(4, 4))

for i, patch in enumerate(patches[0]):

ax = plt.subplot(n, n, i + 1)

patch_img = tf.reshape(patch, (patch_size, patch_size, 3))

plt.imshow(patch_img.numpy().astype("uint8"))

plt.axis("off")

运行结果
Image size: 72 X 72
Patch size: 6 X 6
Patches per image: 144
Elements per patch: 108

7.4构建patch 编码层（ encoding layer）

class PatchEncoder(layers.Layer):
    def __init__(self, num_patches, projection_dim):
        super(PatchEncoder, self).__init__()
        self.num_patches = num_patches
        #一个全连接层，其输出维度为projection_dim，没有指明激活函数
        self.projection = layers.Dense(units=projection_dim)
        #定义一个嵌入层，这是一个可学习的层
        #输入维度为num_patches，输出维度为projection_dim
        self.position_embedding = layers.Embedding(
            input_dim=num_patches, output_dim=projection_dim
        )

    def call(self, patch):
        positions = tf.range(start=0, limit=self.num_patches, delta=1)
        encoded = self.projection(patch) + self.position_embedding(positions)
        return encoded

class PatchEncoder(layers.Layer):

def __init__(self, num_patches, projection_dim):

super(PatchEncoder, self).__init__()

self.num_patches = num_patches

#一个全连接层，其输出维度为projection_dim，没有指明激活函数

self.projection = layers.Dense(units=projection_dim)

#定义一个嵌入层，这是一个可学习的层

#输入维度为num_patches，输出维度为projection_dim

self.position_embedding = layers.Embedding(

input_dim=num_patches, output_dim=projection_dim

)

def call(self, patch):

positions = tf.range(start=0, limit=self.num_patches, delta=1)

encoded = self.projection(patch) + self.position_embedding(positions)

return encoded

7.5构建ViT模型

def create_vit_classifier():
    inputs = layers.Input(shape=input_shape)
    # Augment data.
    augmented = data_augmentation(inputs)
    #augmented = augmented_train_batches(inputs)    
    # Create patches.
    patches = Patches(patch_size)(augmented)
    # Encode patches.
    encoded_patches = PatchEncoder(num_patches, projection_dim)(patches)

    # Create multiple layers of the Transformer block.
    for _ in range(transformer_layers):
        # Layer normalization 1.
        x1 = layers.LayerNormalization(epsilon=1e-6)(encoded_patches)
        # Create a multi-head attention layer.
        attention_output = layers.MultiHeadAttention(
            num_heads=num_heads, key_dim=projection_dim, dropout=0.1
        )(x1, x1)
        # Skip connection 1.
        x2 = layers.Add()([attention_output, encoded_patches])
        # Layer normalization 2.
        x3 = layers.LayerNormalization(epsilon=1e-6)(x2)
        # MLP.
        x3 = mlp(x3, hidden_units=transformer_units, dropout_rate=0.1)
        # Skip connection 2.
        encoded_patches = layers.Add()([x3, x2])

    # Create a [batch_size, projection_dim] tensor.
    representation = layers.LayerNormalization(epsilon=1e-6)(encoded_patches)
    representation = layers.Flatten()(representation)
    representation = layers.Dropout(0.5)(representation)
    # Add MLP.
    features = mlp(representation, hidden_units=mlp_head_units, dropout_rate=0.5)
    # Classify outputs.
    logits = layers.Dense(num_classes)(features)
    # Create the Keras model.
    model = keras.Model(inputs=inputs, outputs=logits)
return model

def create_vit_classifier():

inputs = layers.Input(shape=input_shape)

# Augment data.

augmented = data_augmentation(inputs)

#augmented = augmented_train_batches(inputs)

# Create patches.

patches = Patches(patch_size)(augmented)

# Encode patches.

encoded_patches = PatchEncoder(num_patches, projection_dim)(patches)

# Create multiple layers of the Transformer block.

for _ in range(transformer_layers):

# Layer normalization 1.

x1 = layers.LayerNormalization(epsilon=1e-6)(encoded_patches)

# Create a multi-head attention layer.

attention_output = layers.MultiHeadAttention(

num_heads=num_heads, key_dim=projection_dim, dropout=0.1

)(x1, x1)

# Skip connection 1.

x2 = layers.Add()([attention_output, encoded_patches])

# Layer normalization 2.

x3 = layers.LayerNormalization(epsilon=1e-6)(x2)

# MLP.

x3 = mlp(x3, hidden_units=transformer_units, dropout_rate=0.1)

# Skip connection 2.

encoded_patches = layers.Add()([x3, x2])

# Create a [batch_size, projection_dim] tensor.

representation = layers.LayerNormalization(epsilon=1e-6)(encoded_patches)

representation = layers.Flatten()(representation)

representation = layers.Dropout(0.5)(representation)

# Add MLP.

features = mlp(representation, hidden_units=mlp_head_units, dropout_rate=0.5)

# Classify outputs.

logits = layers.Dense(num_classes)(features)

# Create the Keras model.

model = keras.Model(inputs=inputs, outputs=logits)

return model

该模型的处理流程如下图所示

8、编译、训练模型

def run_experiment(model):
    optimizer = tfa.optimizers.AdamW(
        learning_rate=learning_rate, weight_decay=weight_decay
    )

    model.compile(
        optimizer=optimizer,
        loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True),
        metrics=[
            keras.metrics.SparseCategoricalAccuracy(name="accuracy"),
            keras.metrics.SparseTopKCategoricalAccuracy(5, name="top-5-accuracy"),
        ],
    )

    #checkpoint_filepath = r".\tmp\checkpoint"
    checkpoint_filepath ="model_bak.hdf5"
    checkpoint_callback = keras.callbacks.ModelCheckpoint(
        checkpoint_filepath,
        monitor="val_accuracy",
        save_best_only=True,
        save_weights_only=True,
    )

    history = model.fit(
        x=x_train,
        y=y_train,
        batch_size=batch_size,
        epochs=num_epochs,
        validation_split=0.1,
        callbacks=[checkpoint_callback],
    )

    model.load_weights(checkpoint_filepath)
    _, accuracy, top_5_accuracy = model.evaluate(x_test, y_test)
    print(f"Test accuracy: {round(accuracy * 100, 2)}%")
    print(f"Test top 5 accuracy: {round(top_5_accuracy * 100, 2)}%")

    return history

def run_experiment(model):

optimizer = tfa.optimizers.AdamW(

learning_rate=learning_rate, weight_decay=weight_decay

)

model.compile(

optimizer=optimizer,

loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True),

metrics=[

keras.metrics.SparseCategoricalAccuracy(name="accuracy"),

keras.metrics.SparseTopKCategoricalAccuracy(5, name="top-5-accuracy"),

)

#checkpoint_filepath = r".\tmp\checkpoint"

checkpoint_filepath ="model_bak.hdf5"

checkpoint_callback = keras.callbacks.ModelCheckpoint(

checkpoint_filepath,

monitor="val_accuracy",

save_best_only=True,

save_weights_only=True,

)

history = model.fit(

x=x_train,

y=y_train,

batch_size=batch_size,

epochs=num_epochs,

validation_split=0.1,

callbacks=[checkpoint_callback],

)

model.load_weights(checkpoint_filepath)

_, accuracy, top_5_accuracy = model.evaluate(x_test, y_test)

print(f"Test accuracy: {round(accuracy * 100, 2)}%")

print(f"Test top 5 accuracy: {round(top_5_accuracy * 100, 2)}%")

return history

实例化类，运行模型

vit_classifier = create_vit_classifier()
history = run_experiment(vit_classifier)

1 2	vit_classifier = create_vit_classifier() history = run_experiment(vit_classifier)

运行结果
Epoch 1/10
176/176 [==============================] - 68s 333ms/step - loss: 2.6394 - accuracy: 0.2501 - top-5-accuracy: 0.7377 - val_loss: 1.5331 - val_accuracy: 0.4580 - val_top-5-accuracy: 0.9092
Epoch 2/10
176/176 [==============================] - 58s 327ms/step - loss: 1.6359 - accuracy: 0.4150 - top-5-accuracy: 0.8821 - val_loss: 1.2714 - val_accuracy: 0.5348 - val_top-5-accuracy: 0.9464
Epoch 3/10
176/176 [==============================] - 58s 328ms/step - loss: 1.4332 - accuracy: 0.4839 - top-5-accuracy: 0.9210 - val_loss: 1.1633 - val_accuracy: 0.5806 - val_top-5-accuracy: 0.9616
Epoch 4/10
176/176 [==============================] - 58s 329ms/step - loss: 1.3253 - accuracy: 0.5280 - top-5-accuracy: 0.9349 - val_loss: 1.1010 - val_accuracy: 0.6112 - val_top-5-accuracy: 0.9572
Epoch 5/10
176/176 [==============================] - 58s 330ms/step - loss: 1.2380 - accuracy: 0.5626 - top-5-accuracy: 0.9411 - val_loss: 1.0212 - val_accuracy: 0.6400 - val_top-5-accuracy: 0.9690
Epoch 6/10
176/176 [==============================] - 58s 330ms/step - loss: 1.1486 - accuracy: 0.5945 - top-5-accuracy: 0.9520 - val_loss: 0.9698 - val_accuracy: 0.6602 - val_top-5-accuracy: 0.9718
Epoch 7/10
176/176 [==============================] - 58s 330ms/step - loss: 1.1208 - accuracy: 0.6060 - top-5-accuracy: 0.9558 - val_loss: 0.9215 - val_accuracy: 0.6724 - val_top-5-accuracy: 0.9790
Epoch 8/10
176/176 [==============================] - 58s 330ms/step - loss: 1.0643 - accuracy: 0.6248 - top-5-accuracy: 0.9621 - val_loss: 0.8709 - val_accuracy: 0.6944 - val_top-5-accuracy: 0.9768
Epoch 9/10
176/176 [==============================] - 58s 330ms/step - loss: 1.0119 - accuracy: 0.6446 - top-5-accuracy: 0.9640 - val_loss: 0.8290 - val_accuracy: 0.7142 - val_top-5-accuracy: 0.9784
Epoch 10/10
176/176 [==============================] - 58s 330ms/step - loss: 0.9740 - accuracy: 0.6615 - top-5-accuracy: 0.9666 - val_loss: 0.8175 - val_accuracy: 0.7096 - val_top-5-accuracy: 0.9806
313/313 [==============================] - 9s 27ms/step - loss: 0.8514 - accuracy: 0.7032 - top-5-accuracy: 0.9773
Test accuracy: 70.32%
Test top 5 accuracy: 97.73%
In [15]:
从结果看可以来看，测试精度已达70%，这是一个较大提升！

9、查看运行结果

acc = history.history['accuracy']
val_acc = history.history['val_accuracy']
 
loss = history.history['loss']
val_loss =history.history['val_loss']
 
plt.figure(figsize=(8, 8))
plt.subplot(2, 1, 1)
plt.plot(acc, label='Training Accuracy')
plt.plot(val_acc, label='Validation Accuracy')
plt.legend(loc='lower right')
plt.ylabel('Accuracy')
plt.ylim([min(plt.ylim()),1.1])
plt.title('Training and Validation Accuracy')
 
plt.subplot(2, 1, 2)
plt.plot(loss, label='Training Loss')
plt.plot(val_loss, label='Validation Loss')
plt.legend(loc='upper right')
plt.ylabel('Cross Entropy')
plt.ylim([-0.1,4.0])
plt.title('Training and Validation Loss')
plt.xlabel('epoch')
plt.show()

acc = history.history['accuracy']

val_acc = history.history['val_accuracy']

loss = history.history['loss']

val_loss =history.history['val_loss']

plt.figure(figsize=(8, 8))

plt.subplot(2, 1, 1)

plt.plot(acc, label='Training Accuracy')

plt.plot(val_acc, label='Validation Accuracy')

plt.legend(loc='lower right')

plt.ylabel('Accuracy')

plt.ylim([min(plt.ylim()),1.1])

plt.title('Training and Validation Accuracy')

plt.subplot(2, 1, 2)

plt.plot(loss, label='Training Loss')

plt.plot(val_loss, label='Validation Loss')

plt.legend(loc='upper right')

plt.ylabel('Cross Entropy')

plt.ylim([-0.1,4.0])

plt.title('Training and Validation Loss')

plt.xlabel('epoch')

plt.show()

运行结果

使用数据增强方法提升模型性能

使用数据增强来提升模型的性能

1、导入模型

import os
import math
import numpy as np
import pickle as p
import tensorflow as tf
from tensorflow import keras
import matplotlib.pyplot as plt
%matplotlib inline

import os

import math

import numpy as np

import pickle as p

import tensorflow as tf

from tensorflow import keras

import matplotlib.pyplot as plt

%matplotlib inline

2、定义加载函数

def load_CIFAR_data(data_dir):
    """load CIFAR data"""
 
    images_train=[]
    labels_train=[]
    for i in range(5):
        f=os.path.join(data_dir,'data_batch_%d' % (i+1))
        print('loading ',f)
        # 调用 load_CIFAR_batch( )获得批量的图像及其对应的标签
        image_batch,label_batch=load_CIFAR_batch(f)
        images_train.append(image_batch)
        labels_train.append(label_batch)
        Xtrain=np.concatenate(images_train)
        Ytrain=np.concatenate(labels_train)
        del image_batch ,label_batch
    
    Xtest,Ytest=load_CIFAR_batch(os.path.join(data_dir,'test_batch'))
    print('finished loadding CIFAR-10 data')
    
    # 返回训练集的图像和标签，测试集的图像和标签
return (Xtrain,Ytrain),(Xtest,Ytest)

def load_CIFAR_data(data_dir):

"""load CIFAR data"""

images_train=[]

labels_train=[]

for i in range(5):

f=os.path.join(data_dir,'data_batch_%d' % (i+1))

print('loading ',f)

# 调用 load_CIFAR_batch( )获得批量的图像及其对应的标签

image_batch,label_batch=load_CIFAR_batch(f)

images_train.append(image_batch)

labels_train.append(label_batch)

Xtrain=np.concatenate(images_train)

Ytrain=np.concatenate(labels_train)

del image_batch ,label_batch

Xtest,Ytest=load_CIFAR_batch(os.path.join(data_dir,'test_batch'))

print('finished loadding CIFAR-10 data')

# 返回训练集的图像和标签，测试集的图像和标签

return (Xtrain,Ytrain),(Xtest,Ytest)

3、定义批量加载函数

def load_CIFAR_batch(filename):
    """ load single batch of cifar """  
    with open(filename, 'rb')as f:
        # 一个样本由标签和图像数据组成
        #  (3072=32x32x3)
        # ...
        # 
        data_dict = p.load(f, encoding='bytes')
        images= data_dict[b'data']
        labels = data_dict[b'labels']
                
        # 把原始数据结构调整为: BCWH
        images = images.reshape(10000, 3, 32, 32)
        # tensorflow处理图像数据的结构：BWHC
        # 把通道数据C移动到最后一个维度
        images = images.transpose (0,2,3,1)
     
        labels = np.array(labels)
        
        return images, labels

def load_CIFAR_batch(filename):

""" load single batch of cifar """

with open(filename, 'rb')as f:

# 一个样本由标签和图像数据组成

# (3072=32x32x3)

# ...

data_dict = p.load(f, encoding='bytes')

images= data_dict[b'data']

labels = data_dict[b'labels']

# 把原始数据结构调整为: BCWH

images = images.reshape(10000, 3, 32, 32)

# tensorflow处理图像数据的结构：BWHC

# 把通道数据C移动到最后一个维度

images = images.transpose (0,2,3,1)

labels = np.array(labels)

return images, labels

4、加载数据

data_dir = r'C:\Users\wumg\jupyter-ipynb\data\cifar-10-batches-py'
(x_train,y_train),(x_test,y_test) = load_CIFAR_data(data_dir)

1 2	data_dir = r'C:\Users\wumg\jupyter-ipynb\data\cifar-10-batches-py' (x_train,y_train),(x_test,y_test) = load_CIFAR_data(data_dir)

把数据转换为dataset格式

train_dataset = tf.data.Dataset.from_tensor_slices((x_train, y_train))
test_dataset = tf.data.Dataset.from_tensor_slices((x_test, y_test))

1 2	train_dataset = tf.data.Dataset.from_tensor_slices((x_train, y_train)) test_dataset = tf.data.Dataset.from_tensor_slices((x_test, y_test))

5、定义数据增强方法

def convert(image, label):
    image = tf.image.convert_image_dtype(image, tf.float32)
    return image, label
 
def augment(image, label):
    image, label = convert(image, label)
    #image = tf.image.convert_image_dtype(image, tf.float32)
    image = tf.image.resize_with_crop_or_pad(image, 34,34) # 四周各加3像素
    image = tf.image.random_crop(image, size=[32,32,3]) # 随机裁剪成28*28大小
    image = tf.image.random_brightness(image, max_delta=0.5) # 随机增加亮度
    return image, label
 
batch_size = 64

augmented_train_batches = (train_dataset
                          #.take(num_examples)
                          .cache()
                         # .repeat()
                          .shuffle(5000)
                          .map(augment, num_parallel_calls=tf.data.experimental.AUTOTUNE)
                          .batch(batch_size)
                          .prefetch(tf.data.experimental.AUTOTUNE))
 
non_augmented_train_batches = (train_dataset
                              .cache()
                             # .repeat()
                              .shuffle(5000)
                              .map(convert, num_parallel_calls=tf.data.experimental.AUTOTUNE)
                              .batch(batch_size)
                              .prefetch(tf.data.experimental.AUTOTUNE))
 
validation_batches = (test_dataset 
                     .map(convert, num_parallel_calls=tf.data.experimental.AUTOTUNE)
                     .batch(2*batch_size))

def convert(image, label):

image = tf.image.convert_image_dtype(image, tf.float32)

return image, label

def augment(image, label):

image, label = convert(image, label)

#image = tf.image.convert_image_dtype(image, tf.float32)

image = tf.image.resize_with_crop_or_pad(image, 34,34) # 四周各加3像素

image = tf.image.random_crop(image, size=[32,32,3]) # 随机裁剪成28*28大小

image = tf.image.random_brightness(image, max_delta=0.5) # 随机增加亮度

return image, label

batch_size = 64

augmented_train_batches = (train_dataset

#.take(num_examples)

.cache()

# .repeat()

.shuffle(5000)

.map(augment, num_parallel_calls=tf.data.experimental.AUTOTUNE)

.batch(batch_size)

.prefetch(tf.data.experimental.AUTOTUNE))

non_augmented_train_batches = (train_dataset

.cache()

# .repeat()

.shuffle(5000)

.map(convert, num_parallel_calls=tf.data.experimental.AUTOTUNE)

.batch(batch_size)

.prefetch(tf.data.experimental.AUTOTUNE))

validation_batches = (test_dataset

.map(convert, num_parallel_calls=tf.data.experimental.AUTOTUNE)

.batch(2*batch_size))

6、构建模型

class MyCNN(tf.keras.Model):
    def __init__(self):
        super().__init__()
        self.conv1 = tf.keras.layers.Conv2D(
            filters=32,             # 卷积层神经元（卷积核）数目
            kernel_size=[3, 3],     # 感受野大小
            padding='same',         # padding策略（vaild 或 same）
            activation=tf.nn.relu   # 激活函数
        )
        self.pool1 = tf.keras.layers.MaxPool2D(pool_size=[2, 2], strides=2)
        self.conv2 = tf.keras.layers.Conv2D(
            filters=64,
            kernel_size=[3, 3],
            padding='same',
            activation=tf.nn.relu
        )
        self.pool2 = tf.keras.layers.MaxPool2D(pool_size=[2, 2], strides=2)
        self.flatten = tf.keras.layers.Reshape(target_shape=(8 * 8 * 64,))
        self.dense1 = tf.keras.layers.Dense(units=256, activation=tf.nn.relu)
        self.dense2 = tf.keras.layers.Dense(units=10)

    def call(self, inputs):
        x = self.conv1(inputs)                  # [batch_size, 32, 32, 3]
        x = self.pool1(x)                       # [batch_size, 32, 32, 32]
        x = self.conv2(x)                       # [batch_size, 16, 16, 64]
        x = self.pool2(x)                       # [batch_size, 8, 8, 64]
        x = self.flatten(x)                     # [batch_size, 8 * 8 * 64]
        x = self.dense1(x)                      # [batch_size, 256]
        x = self.dense2(x)                      # [batch_size, 10]
        output = tf.nn.softmax(x)
        return output
    def model01(self):
        x = tf.keras.Input(shape=(32, 32, 3))
        return tf.keras.Model(inputs=[x], outputs=self.call(x))

class MyCNN(tf.keras.Model):

def __init__(self):

super().__init__()

self.conv1 = tf.keras.layers.Conv2D(

filters=32, # 卷积层神经元（卷积核）数目

kernel_size=[3, 3], # 感受野大小

padding='same', # padding策略（vaild 或 same）

activation=tf.nn.relu # 激活函数

)

self.pool1 = tf.keras.layers.MaxPool2D(pool_size=[2, 2], strides=2)

self.conv2 = tf.keras.layers.Conv2D(

filters=64,

kernel_size=[3, 3],

padding='same',

activation=tf.nn.relu

)

self.pool2 = tf.keras.layers.MaxPool2D(pool_size=[2, 2], strides=2)

self.flatten = tf.keras.layers.Reshape(target_shape=(8 * 8 * 64,))

self.dense1 = tf.keras.layers.Dense(units=256, activation=tf.nn.relu)

self.dense2 = tf.keras.layers.Dense(units=10)

def call(self, inputs):

x = self.conv1(inputs) # [batch_size, 32, 32, 3]

x = self.pool1(x) # [batch_size, 32, 32, 32]

x = self.conv2(x) # [batch_size, 16, 16, 64]

x = self.pool2(x) # [batch_size, 8, 8, 64]

x = self.flatten(x) # [batch_size, 8 * 8 * 64]

x = self.dense1(x) # [batch_size, 256]

x = self.dense2(x) # [batch_size, 10]

output = tf.nn.softmax(x)

return output

def model01(self):

x = tf.keras.Input(shape=(32, 32, 3))

return tf.keras.Model(inputs=[x], outputs=self.call(x))

生成实例

model_no_augment = MyCNN()

1	model_no_augment = MyCNN()

查看模型详细结构

model_no_augment.model01().summary()

1	model_no_augment.model01().summary()

7、编译模型

model_no_augment.compile(optimizer='adam',loss='sparse_categorical_crossentropy',metrics=['accuracy'])

1	model_no_augment.compile(optimizer='adam',loss='sparse_categorical_crossentropy',metrics=['accuracy'])

8、训练模型

为便于比较，这里先不使用数据增强方法

epochs = 10
history_non_augment = model_no_augment.fit(non_augmented_train_batches,epochs=epochs,validation_data=validation_batches)

1 2	epochs = 10 history_non_augment = model_no_augment.fit(non_augmented_train_batches,epochs=epochs,validation_data=validation_batches)

9、查看运行结果

acc = history_non_augment.history['accuracy']
val_acc = history_non_augment.history['val_accuracy']
 
loss = history_non_augment.history['loss']
val_loss = history_non_augment.history['val_loss']
 
plt.figure(figsize=(8, 8))
plt.subplot(2, 1, 1)
plt.plot(acc, label='Training Accuracy')
plt.plot(val_acc, label='Validation Accuracy')
plt.legend(loc='lower right')
plt.ylabel('Accuracy')
plt.ylim([min(plt.ylim()),1.1])
plt.title('Training and Validation Accuracy')
 
plt.subplot(2, 1, 2)
plt.plot(loss, label='Training Loss')
plt.plot(val_loss, label='Validation Loss')
plt.legend(loc='upper right')
plt.ylabel('Cross Entropy')
plt.ylim([-0.1,1.0])
plt.title('Training and Validation Loss')
plt.xlabel('epoch')
plt.show()

acc = history_non_augment.history['accuracy']

val_acc = history_non_augment.history['val_accuracy']

loss = history_non_augment.history['loss']

val_loss = history_non_augment.history['val_loss']

plt.figure(figsize=(8, 8))

plt.subplot(2, 1, 1)

plt.plot(acc, label='Training Accuracy')

plt.plot(val_acc, label='Validation Accuracy')

plt.legend(loc='lower right')

plt.ylabel('Accuracy')

plt.ylim([min(plt.ylim()),1.1])

plt.title('Training and Validation Accuracy')

plt.subplot(2, 1, 2)

plt.plot(loss, label='Training Loss')

plt.plot(val_loss, label='Validation Loss')

plt.legend(loc='upper right')

plt.ylabel('Cross Entropy')

plt.ylim([-0.1,1.0])

plt.title('Training and Validation Loss')

plt.xlabel('epoch')

plt.show()

运行结果

10、使用数据增强方法

model_augment = MyCNN()

model_augment.compile(optimizer='adam',loss='sparse_categorical_crossentropy',metrics=['accuracy'])

history_with_augment = model_augment.fit(augmented_train_batches,epochs=epochs,validation_data=validation_batches)

model_augment = MyCNN()

model_augment.compile(optimizer='adam',loss='sparse_categorical_crossentropy',metrics=['accuracy'])

history_with_augment = model_augment.fit(augmented_train_batches,epochs=epochs,validation_data=validation_batches)

11、查看使用数据增强的运行结果

acc = history_with_augment.history['accuracy']
val_acc = history_with_augment.history['val_accuracy']
 
loss = history_with_augment.history['loss']
val_loss = history_with_augment.history['val_loss']
 
plt.figure(figsize=(8, 8))
plt.subplot(2, 1, 1)
plt.plot(acc, label='Training Accuracy')
plt.plot(val_acc, label='Validation Accuracy')
plt.legend(loc='lower right')
plt.ylabel('Accuracy')
plt.ylim([min(plt.ylim()),1.1])
plt.title('Training and Validation Accuracy')
 
plt.subplot(2, 1, 2)
plt.plot(loss, label='Training Loss')
plt.plot(val_loss, label='Validation Loss')
plt.legend(loc='upper right')
plt.ylabel('Cross Entropy')
plt.ylim([-0.1,1.0])
plt.title('Training and Validation Loss')
plt.xlabel('epoch')
plt.show()

acc = history_with_augment.history['accuracy']

val_acc = history_with_augment.history['val_accuracy']

loss = history_with_augment.history['loss']

val_loss = history_with_augment.history['val_loss']

plt.figure(figsize=(8, 8))

plt.subplot(2, 1, 1)

plt.plot(acc, label='Training Accuracy')

plt.plot(val_acc, label='Validation Accuracy')

plt.legend(loc='lower right')

plt.ylabel('Accuracy')

plt.ylim([min(plt.ylim()),1.1])

plt.title('Training and Validation Accuracy')

plt.subplot(2, 1, 2)

plt.plot(loss, label='Training Loss')

plt.plot(val_loss, label='Validation Loss')

plt.legend(loc='upper right')

plt.ylabel('Cross Entropy')

plt.ylim([-0.1,1.0])

plt.title('Training and Validation Loss')

plt.xlabel('epoch')

plt.show()

运行结果

12、结果分析

从不使用数据增强与使用数据增强方法的结果可以看出，使用数据增强方法后，模型性能有提升(未使用数据增强的验证精度为71%，使用数据增强方法后，验证精度提升到74%)，而且模型的泛化能力也有提高（使用数据增强方法后，训练与验证精度曲线靠得较近）。

TensorFlow2提升模型性能的几种有效方法

这里介绍几种TensorFlow2提升模型性能的几种有效方法，如数据增强、使用经典网络、使用迁移学习、使用新架构等。

一、数据增强

数据增强是提升模型性能的常用方法之一，尤其当训练数据集较小时，通过数据增强方法，效果更加明确。通过数据增强不但可以增加数据量、丰富数据的多样性，从而有效提升模型的泛化能力。
TensorFlow2提供了多种数据增强方法，常用的几种方法为：
 有监督的数据增强：
1、使用tf.image的预处理
该方法一般基于dataset，然后把数据增强方法嵌入到map中
2、使用tf.keras.preprocessing.image.ImageDataGenerator
利用实时数据增强技术生成批量的张量图像数据
使用tf.keras.preprocessing.image.ImageDataGenerator，获取数据源的方法主要有：
（1）.flow(x, y)
（2）.flow_from_directory(directory)
3、使用tf.keras.layers.experimental.preprocessing
更多信息可参考官网：https://keras.io/guides/preprocessing_layers/
 无监督的数据增强
通过模型学习数据的分布，随机生成与训练数据集分布一致的图片，代表方法，GAN。
GAN模型包含两个网络，一个是生成网络（G），一个是判别网络（D），基本原理如下：
(1)G是一个生成图片的网络，它接收随机的噪声z，通过噪声生成图片，记做G(z)。
(2)D是一个判别网络，判别一张图片是不是“真实的”，即是真实的图片，还是由G生成的图片。其架构图如下：

下载所有实例使用数据（提取码为：fg29）
具体实例

二、使用现代经典模型

现代经典模型主要有: VGG、GoogLeNet、Inception、Xception、ResNet、MobileNet、DenseNet、NASNet等,各种网咯结构可参考：
https://blog.csdn.net/Forrest97/article/details/105630719

具体实例(待续)

三、利用迁移方法

Tensorflow.keras.application下载已经训练好的模型，包括如下预训练模型：
VGG、GoogLeNet、Inception、Xception、ResNet、MobileNet、DenseNet、NASNet等

具体实例(待续)

四、使用新架构如Transformer等

Transformer）模型在NLP领域取得了SOTA成绩，目前人们正把Transformer引用到CV领域，在图像识别、图像分类等领域取得不俗的表现。在目标检测（如DETR）、图像分类（如ViT）、图像分割（如SETR）都取得不错的效果。这里我们重点介绍ViT（Vision Transformer）。如何把Transformer引入CV领域？需要对CV中的图像做哪些处理？
NLP处理的语言数据是序列化的，而CV中处理的图像数据是三维的（height、width和channels）。所以需要通过某种方法将图像这种三维数据转化为序列化的数据。
Vision Transformer将CV和NLP领域知识结合起来，对原始图片进行分块，然后展平成序列，输入进原始Transformer模型的编码器Encoder部分，最后接入一个全连接层对图片进行分类。具体步骤为：
（1）原始图片（H,W,C）进行分块（patches）
进行一个类似卷积操作，把原始图片进行分块，分块数=H*W/P*P（P为块的大小）。
（2）展平每块（Flatten the patches）
把每块展平为一维向量，大小为：P*P*C（对应实例中的patch_dims）
（3）把展平后的块映射为更低维的向量
通过一个全连接层，把展平后块映射为一个更低维的向量(对应实例中的patch embedding)，其大小为D(对应实例中的projection_dim),这个维度在各层保存不变，主要为便于使用残差连接。
（4）在patch embedding基础上添加一个类标签
类似BERT的[class] token，在patch embedding的序列之前添加一个可学习的embedding向量xclass
（5）加上位置嵌入（Add positional embeddings）
低维的向量+位置嵌入,
（6）把（5）的结果作为标准transformer encoder的输入
（7）在一个大数据集上训练模型
（8）在下游数据集中，进行微调。
ViT架构图如下：

具体实例

五、练习

1、利用ResNet经典模型提升性能
2、使用keras的数据增强方法提升性能
3、viT中Transformer的输入数据的形状（shape）与哪些因素有关？包括哪些数据？如何生成Transformer的输入数据？

TensorFlow2构建模型-境界4

1.导入CIFAR-10数据集

CIFAR-10是由 Hinton 的学生 Alex Krizhevsky 和 Ilya Sutskever 整理的一个用于识别普适物体的小型数据集。一共包含 10 个类别的 RGB 彩色图片：飞机（ a叩lane ）、汽车（ automobile ）、鸟类（ bird ）、猫（ cat ）、鹿（ deer ）、狗（ dog ）、蛙类（ frog ）、马（ horse ）、船（ ship ）和卡车（ truck ）。图片的尺寸为 32×32，3个通道，数据集中一共有 50000 张训练圄片和 10000 张测试图片。 CIFAR-10数据集有3个版本，这里使用python版本。

1.1 导入需要的库

import os
import math
import numpy as np
import pickle as p
import tensorflow as tf
import matplotlib.pyplot as plt
%matplotlib inline

import os

import math

import numpy as np

import pickle as p

import tensorflow as tf

import matplotlib.pyplot as plt

%matplotlib inline

1.2 定义批量导入数据的函数

def load_CIFAR_batch(filename):
    """ load single batch of cifar """  
    with open(filename, 'rb')as f:
        # 一个样本由标签和图像数据组成
        #  (3072=32x32x3)
        # ...
        # 
        data_dict = p.load(f, encoding='bytes')
        images= data_dict[b'data']
        labels = data_dict[b'labels']
                
        # 把原始数据结构调整为: BCWH
        images = images.reshape(10000, 3, 32, 32)
        # tensorflow处理图像数据的结构：BWHC
        # 把通道数据C移动到最后一个维度
        images = images.transpose (0,2,3,1)
     
        labels = np.array(labels)
        
        return images, labels

def load_CIFAR_batch(filename):

""" load single batch of cifar """

with open(filename, 'rb')as f:

# 一个样本由标签和图像数据组成

# (3072=32x32x3)

# ...

data_dict = p.load(f, encoding='bytes')

images= data_dict[b'data']

labels = data_dict[b'labels']

# 把原始数据结构调整为: BCWH

images = images.reshape(10000, 3, 32, 32)

# tensorflow处理图像数据的结构：BWHC

# 把通道数据C移动到最后一个维度

images = images.transpose (0,2,3,1)

labels = np.array(labels)

return images, labels

1.3 定义加载数据函数

def load_CIFAR_data(data_dir):
    """load CIFAR data"""
 
    images_train=[]
    labels_train=[]
    for i in range(5):
        f=os.path.join(data_dir,'data_batch_%d' % (i+1))
        print('loading ',f)
        # 调用 load_CIFAR_batch( )获得批量的图像及其对应的标签
        image_batch,label_batch=load_CIFAR_batch(f)
        images_train.append(image_batch)
        labels_train.append(label_batch)
        Xtrain=np.concatenate(images_train)
        Ytrain=np.concatenate(labels_train)
        del image_batch ,label_batch
    
    Xtest,Ytest=load_CIFAR_batch(os.path.join(data_dir,'test_batch'))
    print('finished loadding CIFAR-10 data')
    
    # 返回训练集的图像和标签，测试集的图像和标签
    return (Xtrain,Ytrain),(Xtest,Ytest)

def load_CIFAR_data(data_dir):

"""load CIFAR data"""

images_train=[]

labels_train=[]

for i in range(5):

f=os.path.join(data_dir,'data_batch_%d' % (i+1))

print('loading ',f)

# 调用 load_CIFAR_batch( )获得批量的图像及其对应的标签

image_batch,label_batch=load_CIFAR_batch(f)

images_train.append(image_batch)

labels_train.append(label_batch)

Xtrain=np.concatenate(images_train)

Ytrain=np.concatenate(labels_train)

del image_batch ,label_batch

Xtest,Ytest=load_CIFAR_batch(os.path.join(data_dir,'test_batch'))

print('finished loadding CIFAR-10 data')

# 返回训练集的图像和标签，测试集的图像和标签

return (Xtrain,Ytrain),(Xtest,Ytest)

1.4 加载数据

data_dir = r'data\cifar-10-batches-py'
(x_train,y_train),(x_test,y_test) = load_CIFAR_data(data_dir)

1 2	data_dir = r'data\cifar-10-batches-py' (x_train,y_train),(x_test,y_test) = load_CIFAR_data(data_dir)

运行结果
loading data\cifar-10-batches-py\data_batch_1
loading data\cifar-10-batches-py\data_batch_2
loading data\cifar-10-batches-py\data_batch_3
loading data\cifar-10-batches-py\data_batch_4
loading data\cifar-10-batches-py\data_batch_5
finished loadding CIFAR-10 data

1.5 可视化加载数据

label_dict = {0:"airplane", 1:"
automobile", 2:"bird", 3:"cat", 4:"deer",
              5:"dog", 6:"frog", 7:"horse", 8:"ship", 9:"truck"}


def plot_images_labels(images, labels, num):
    total = len(images)
    fig = plt.gcf()
    fig.set_size_inches(15, math.ceil(num / 10) * 7)
    for i in range(0, num):
        choose_n = np.random.randint(0, total)
        
        ax = plt.subplot(math.ceil(num / 5), 5, 1 + i)
        ax.imshow(images[choose_n], cmap='binary')
        title = label_dict[labels[choose_n]]
        ax.set_title(title, fontsize=10)        
    plt.show()
	
plot_images_labels(x_train, y_train, 10)

label_dict = {0:"airplane", 1:"

automobile", 2:"bird", 3:"cat", 4:"deer",

5:"dog", 6:"frog", 7:"horse", 8:"ship", 9:"truck"}

def plot_images_labels(images, labels, num):

total = len(images)

fig = plt.gcf()

fig.set_size_inches(15, math.ceil(num / 10) * 7)

for i in range(0, num):

choose_n = np.random.randint(0, total)

ax = plt.subplot(math.ceil(num / 5), 5, 1 + i)

ax.imshow(images[choose_n], cmap='binary')

title = label_dict[labels[choose_n]]

ax.set_title(title, fontsize=10)

plt.show()

plot_images_labels(x_train, y_train, 10)

运行结果

2 .数据预处理并设置超参数

x_train = x_train.astype('float32') / 255.0
x_test = x_test.astype('float32') / 255.0

train_num = len(x_train)
num_classes = 10

learning_rate = 0.0002
batch_size = 64
training_steps = 20000
display_step = 1000

conv1_filters = 32
conv2_filters = 64
fc1_units = 256

x_train = x_train.astype('float32') / 255.0

x_test = x_test.astype('float32') / 255.0

train_num = len(x_train)

num_classes = 10

learning_rate = 0.0002

batch_size = 64

training_steps = 20000

display_step = 1000

conv1_filters = 32

conv2_filters = 64

fc1_units = 256

3.使用tf.data构建数据管道

AUTOTUNE = tf.data.experimental.AUTOTUNE
train_data = tf.data.Dataset.from_tensor_slices((x_train, y_train))
train_data = train_data.shuffle(5000).repeat(training_steps).batch(batch_size).prefetch(buffer_size=AUTOTUNE)

AUTOTUNE = tf.data.experimental.AUTOTUNE

train_data = tf.data.Dataset.from_tensor_slices((x_train, y_train))

train_data = train_data.shuffle(5000).repeat(training_steps).batch(batch_size).prefetch(buffer_size=AUTOTUNE)

4.构建模型

使用子类方法自定义模型结构的一般步骤：

class MyCNN(tf.keras.Model):
    def __init__(self):
        super().__init__()
        self.conv1 = tf.keras.layers.Conv2D(
            filters=32,             # 卷积层神经元（卷积核）数目
            kernel_size=[3, 3],     # 感受野大小
            padding='same',         # padding策略（vaild 或 same）
            activation=tf.nn.relu   # 激活函数
        )
        self.pool1 = tf.keras.layers.MaxPool2D(pool_size=[2, 2], strides=2)
        self.conv2 = tf.keras.layers.Conv2D(
            filters=64,
            kernel_size=[3, 3],
            padding='same',
            activation=tf.nn.relu
        )
        self.pool2 = tf.keras.layers.MaxPool2D(pool_size=[2, 2], strides=2)
        self.flatten = tf.keras.layers.Reshape(target_shape=(8 * 8 * 64,))
        self.dense1 = tf.keras.layers.Dense(units=256, activation=tf.nn.relu)
        self.dense2 = tf.keras.layers.Dense(units=10)

  
    def call(self, inputs):
        x = self.conv1(inputs)                  # [batch_size, 32, 32, 3]
        x = self.pool1(x)                       # [batch_size, 32, 32, 32]
        x = self.conv2(x)                       # [batch_size, 16, 16, 64]
        x = self.pool2(x)                       # [batch_size, 8, 8, 64]
        x = self.flatten(x)                     # [batch_size, 8 * 8 * 64]
        x = self.dense1(x)                      # [batch_size, 256]
        x = self.dense2(x)                      # [batch_size, 10]
        output = tf.nn.softmax(x)
        return output
    #为使用summary时能显示tensor的shape
    def model01(self):
        x = tf.keras.Input(shape=(32, 32, 3))
        return tf.keras.Model(inputs=[x], outputs=self.call(x))

class MyCNN(tf.keras.Model):

def __init__(self):

super().__init__()

self.conv1 = tf.keras.layers.Conv2D(

filters=32, # 卷积层神经元（卷积核）数目

kernel_size=[3, 3], # 感受野大小

padding='same', # padding策略（vaild 或 same）

activation=tf.nn.relu # 激活函数

)

self.pool1 = tf.keras.layers.MaxPool2D(pool_size=[2, 2], strides=2)

self.conv2 = tf.keras.layers.Conv2D(

filters=64,

kernel_size=[3, 3],

padding='same',

activation=tf.nn.relu

)

self.pool2 = tf.keras.layers.MaxPool2D(pool_size=[2, 2], strides=2)

self.flatten = tf.keras.layers.Reshape(target_shape=(8 * 8 * 64,))

self.dense1 = tf.keras.layers.Dense(units=256, activation=tf.nn.relu)

self.dense2 = tf.keras.layers.Dense(units=10)

def call(self, inputs):

x = self.conv1(inputs) # [batch_size, 32, 32, 3]

x = self.pool1(x) # [batch_size, 32, 32, 32]

x = self.conv2(x) # [batch_size, 16, 16, 64]

x = self.pool2(x) # [batch_size, 8, 8, 64]

x = self.flatten(x) # [batch_size, 8 * 8 * 64]

x = self.dense1(x) # [batch_size, 256]

x = self.dense2(x) # [batch_size, 10]

output = tf.nn.softmax(x)

return output

#为使用summary时能显示tensor的shape

def model01(self):

x = tf.keras.Input(shape=(32, 32, 3))

return tf.keras.Model(inputs=[x], outputs=self.call(x))

5.训练模型

（1）实例化模型
model = MyCNN()
（2）查看模型的详细结构
model.model01().summary()
Model: "model"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
input_1 (InputLayer) [(None, 32, 32, 3)] 0
_________________________________________________________________
conv2d (Conv2D) (None, 32, 32, 32) 896
_________________________________________________________________
max_pooling2d (MaxPooling2D) (None, 16, 16, 32) 0
_________________________________________________________________
conv2d_1 (Conv2D) (None, 16, 16, 64) 18496
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 8, 8, 64) 0
_________________________________________________________________
reshape (Reshape) (None, 4096) 0
_________________________________________________________________
dense (Dense) (None, 256) 1048832
_________________________________________________________________
dense_1 (Dense) (None, 10) 2570
_________________________________________________________________
tf_op_layer_Softmax (TensorF [(None, 10)] 0
=================================================================
Total params: 1,070,794
Trainable params: 1,070,794
Non-trainable params: 0
（3）编译及训练模型
model.compile(optimizer='adam',loss='sparse_categorical_crossentropy',metrics=['accuracy'])
epochs = 10
batch_size = 64
learning_rate = 0.0002

train_history = model.fit(x_train, y_train,
validation_split=0.2,
epochs=epochs,
#steps_per_epoch=100,
batch_size=batch_size,
verbose=1)
运行结果
Train on 40000 samples, validate on 10000 samples
Epoch 1/10
40000/40000 [==============================] - 5s 113us/sample - loss: 1.4482 - accuracy: 0.4811 - val_loss: 1.1808 - val_accuracy: 0.5889
Epoch 2/10
40000/40000 [==============================] - 2s 53us/sample - loss: 1.0530 - accuracy: 0.6291 - val_loss: 1.0052 - val_accuracy: 0.6466
Epoch 3/10
40000/40000 [==============================] - 2s 51us/sample - loss: 0.9031 - accuracy: 0.6814 - val_loss: 0.9358 - val_accuracy: 0.6751
Epoch 4/10
40000/40000 [==============================] - 2s 50us/sample - loss: 0.7926 - accuracy: 0.7207 - val_loss: 0.8919 - val_accuracy: 0.6909
Epoch 5/10
40000/40000 [==============================] - 2s 53us/sample - loss: 0.6966 - accuracy: 0.7573 - val_loss: 0.8932 - val_accuracy: 0.6904
Epoch 6/10
40000/40000 [==============================] - 2s 50us/sample - loss: 0.6029 - accuracy: 0.7889 - val_loss: 0.8699 - val_accuracy: 0.7036
Epoch 7/10
40000/40000 [==============================] - 2s 51us/sample - loss: 0.5131 - accuracy: 0.8210 - val_loss: 0.8832 - val_accuracy: 0.7092
Epoch 8/10
40000/40000 [==============================] - 2s 52us/sample - loss: 0.4263 - accuracy: 0.8533 - val_loss: 0.9517 - val_accuracy: 0.7028
Epoch 9/10
40000/40000 [==============================] - 2s 53us/sample - loss: 0.3407 - accuracy: 0.8815 - val_loss: 0.9970 - val_accuracy: 0.7065
Epoch 10/10
40000/40000 [==============================] - 2s 52us/sample - loss: 0.2693 - accuracy: 0.9078 - val_loss: 1.0540 - val_accuracy: 0.7090

6.可视化运行结果

plt.title('the train and validate')
plt.xlabel('Times')
plt.ylabel('Loss value')
plt.plot(train_history.history['loss'], color=(1, 0, 0), label='Loss train')
plt.plot(train_history.history['val_accuracy'], color=(0, 0, 1), label='Accuracy train')
plt.legend(loc='best')
plt.show()

plt.title('the train and validate')

plt.xlabel('Times')

plt.ylabel('Loss value')

plt.plot(train_history.history['loss'], color=(1, 0, 0), label='Loss train')

plt.plot(train_history.history['val_accuracy'], color=(0, 0, 1), label='Accuracy train')

plt.legend(loc='best')

plt.show()

7.测试模型

test_loss, test_acc = model.evaluate(x_test, y_test, verbose=2)
print('test_loss:', test_loss,'\ntest_acc:', test_acc,'\nmetrics_names:', model.metrics_names)

1 2	test_loss, test_acc = model.evaluate(x_test, y_test, verbose=2) print('test_loss:', test_loss,'\ntest_acc:', test_acc,'\nmetrics_names:', model.metrics_names)

运行结果
10000/10000 - 1s - loss: 1.1505 - accuracy: 0.6936
test_loss: 1.1505295364379884
test_acc: 0.6936
metrics_names: ['loss', 'accuracy']

8.保存恢复整个模型

（1）保存模型参数及网络结构等
可以使用两种格式将整个模型保存到磁盘：TensorFlow SavedModel 格式和较早的 Keras H5 格式。 tensorflow官方推荐使用 SavedModel 格式。它是使用 model.save() 时的默认格式,这种保存方式适合Sequential, Functional Model, or Model subclass。

model.save('my_model')

1	model.save('my_model')

（2）恢复模型

newmodel = keras.models.load_model('my_model')

1	newmodel = keras.models.load_model('my_model')

（3）检查恢复模型的结构

newmodel.summary()

1	newmodel.summary()

运行结果
Model: "my_cnn_1"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
conv2d_2 (Conv2D) multiple 896
_________________________________________________________________
max_pooling2d_2 (MaxPooling2 multiple 0
_________________________________________________________________
conv2d_3 (Conv2D) multiple 18496
_________________________________________________________________
max_pooling2d_3 (MaxPooling2 multiple 0
_________________________________________________________________
reshape_1 (Reshape) multiple 0
_________________________________________________________________
dense_2 (Dense) multiple 1048832
_________________________________________________________________
dense_3 (Dense) multiple 2570
=================================================================
Total params: 1,070,794
Trainable params: 1,070,794
Non-trainable params: 0
_______________________________
与原模型结构完全一致！

（4）基于恢复模型进行测试

test_loss, test_acc = newmodel.evaluate(x_test, y_test, verbose=2)
print('test_loss:', test_loss,'\ntest_acc:', test_acc,'\nmetrics_names:', model.metrics_names)

1 2	test_loss, test_acc = newmodel.evaluate(x_test, y_test, verbose=2) print('test_loss:', test_loss,'\ntest_acc:', test_acc,'\nmetrics_names:', model.metrics_names)

运行结果
10000/10000 - 1s - loss: 1.1488 - accuracy: 0.6936
test_loss: 1.1488204129219055
test_acc: 0.6936
metrics_names: ['loss', 'accuracy']
模型精度与原模型完全一致！

TensorFlow2 构建模型-境界3

有三种计算图的构建方式：静态计算图，动态计算图，以及Autograph。 TensorFlow 2.0主要使用的是动态计算图和Autograph。动态计算图易于调试，编码效率较高，但执行效率偏低。静态计算图执行效率很高，但较难调试。而Autograph机制可以将动态图转换成静态计算图，兼收执行效率和编码效率之利。当然Autograph机制能够转换的代码并不是没有任何约束的，有一些编码规范需要遵循，否则可能会转换失败或者不符合预期。我们将着重介绍Autograph的编码规范和Autograph转换成静态图的原理。
Autograph编码规范:
• 1.被@tf.function修饰的函数应尽可能使用TensorFlow中的函数而不是Python中的其他函数。例如使用tf.print而不是print，使用tf.range而不是range，使用tf.constant(True)而不是True.
• 2.避免在@tf.function修饰的函数内部定义tf.Variable.
• 3.被@tf.function修饰的函数不可修改该函数外部的Python列表或字典等数据结构变量。

1.导入CIFAR-10数据集

1.1 导入需要的库

import os
import math
import numpy as np
import pickle as p
import tensorflow as tf
import matplotlib.pyplot as plt
%matplotlib inline

import os

import math

import numpy as np

import pickle as p

import tensorflow as tf

import matplotlib.pyplot as plt

%matplotlib inline

1.2 定义批量导入数据的函数

def load_CIFAR_batch(filename):
    """ load single batch of cifar """  
    with open(filename, 'rb')as f:
        # 一个样本由标签和图像数据组成
        #  (3072=32x32x3)
        # ...
        # 
        data_dict = p.load(f, encoding='bytes')
        images= data_dict[b'data']
        labels = data_dict[b'labels']
                
        # 把原始数据结构调整为: BCWH
        images = images.reshape(10000, 3, 32, 32)
        # tensorflow处理图像数据的结构：BWHC
        # 把通道数据C移动到最后一个维度
        images = images.transpose (0,2,3,1)
     
        labels = np.array(labels)
        
        return images, labels

def load_CIFAR_batch(filename):

""" load single batch of cifar """

with open(filename, 'rb')as f:

# 一个样本由标签和图像数据组成

# (3072=32x32x3)

# ...

data_dict = p.load(f, encoding='bytes')

images= data_dict[b'data']

labels = data_dict[b'labels']

# 把原始数据结构调整为: BCWH

images = images.reshape(10000, 3, 32, 32)

# tensorflow处理图像数据的结构：BWHC

# 把通道数据C移动到最后一个维度

images = images.transpose (0,2,3,1)

labels = np.array(labels)

return images, labels

1.3 定义加载数据函数

def load_CIFAR_data(data_dir):
    """load CIFAR data"""
 
    images_train=[]
    labels_train=[]
    for i in range(5):
        f=os.path.join(data_dir,'data_batch_%d' % (i+1))
        print('loading ',f)
        # 调用 load_CIFAR_batch( )获得批量的图像及其对应的标签
        image_batch,label_batch=load_CIFAR_batch(f)
        images_train.append(image_batch)
        labels_train.append(label_batch)
        Xtrain=np.concatenate(images_train)
        Ytrain=np.concatenate(labels_train)
        del image_batch ,label_batch
    
    Xtest,Ytest=load_CIFAR_batch(os.path.join(data_dir,'test_batch'))
    print('finished loadding CIFAR-10 data')
    
    # 返回训练集的图像和标签，测试集的图像和标签
    return (Xtrain,Ytrain),(Xtest,Ytest)

def load_CIFAR_data(data_dir):

"""load CIFAR data"""

images_train=[]

labels_train=[]

for i in range(5):

f=os.path.join(data_dir,'data_batch_%d' % (i+1))

print('loading ',f)

# 调用 load_CIFAR_batch( )获得批量的图像及其对应的标签

image_batch,label_batch=load_CIFAR_batch(f)

images_train.append(image_batch)

labels_train.append(label_batch)

Xtrain=np.concatenate(images_train)

Ytrain=np.concatenate(labels_train)

del image_batch ,label_batch

Xtest,Ytest=load_CIFAR_batch(os.path.join(data_dir,'test_batch'))

print('finished loadding CIFAR-10 data')

# 返回训练集的图像和标签，测试集的图像和标签

return (Xtrain,Ytrain),(Xtest,Ytest)

1.4 加载数据

data_dir = r'data\cifar-10-batches-py'
(x_train,y_train),(x_test,y_test) = load_CIFAR_data(data_dir)

1 2	data_dir = r'data\cifar-10-batches-py' (x_train,y_train),(x_test,y_test) = load_CIFAR_data(data_dir)

1.5 可视化加载数据

label_dict = {0:"airplane", 1:"
automobile", 2:"bird", 3:"cat", 4:"deer",
              5:"dog", 6:"frog", 7:"horse", 8:"ship", 9:"truck"}


def plot_images_labels(images, labels, num):
    total = len(images)
    fig = plt.gcf()
    fig.set_size_inches(15, math.ceil(num / 10) * 7)
    for i in range(0, num):
        choose_n = np.random.randint(0, total)
        
        ax = plt.subplot(math.ceil(num / 5), 5, 1 + i)
        ax.imshow(images[choose_n], cmap='binary')
        title = label_dict[labels[choose_n]]
        ax.set_title(title, fontsize=10)        
    plt.show()
	
plot_images_labels(x_train, y_train, 10)

label_dict = {0:"airplane", 1:"

automobile", 2:"bird", 3:"cat", 4:"deer",

5:"dog", 6:"frog", 7:"horse", 8:"ship", 9:"truck"}

def plot_images_labels(images, labels, num):

total = len(images)

fig = plt.gcf()

fig.set_size_inches(15, math.ceil(num / 10) * 7)

for i in range(0, num):

choose_n = np.random.randint(0, total)

ax = plt.subplot(math.ceil(num / 5), 5, 1 + i)

ax.imshow(images[choose_n], cmap='binary')

title = label_dict[labels[choose_n]]

ax.set_title(title, fontsize=10)

plt.show()

plot_images_labels(x_train, y_train, 10)

运行结果

2 .数据预处理并设置超参数

x_train = x_train.astype('float32') / 255.0
x_test = x_test.astype('float32') / 255.0

train_num = len(x_train)
num_classes = 10

learning_rate = 0.0002
batch_size = 64
training_steps = 20000
display_step = 1000

conv1_filters = 32
conv2_filters = 64
fc1_units = 256

x_train = x_train.astype('float32') / 255.0

x_test = x_test.astype('float32') / 255.0

train_num = len(x_train)

num_classes = 10

learning_rate = 0.0002

batch_size = 64

training_steps = 20000

display_step = 1000

conv1_filters = 32

conv2_filters = 64

fc1_units = 256

3.使用tf,data构建数据管道

AUTOTUNE = tf.data.experimental.AUTOTUNE
train_data = tf.data.Dataset.from_tensor_slices((x_train, y_train))
train_data = train_data.shuffle(5000).repeat(training_steps).batch(batch_size).prefetch(buffer_size=AUTOTUNE)

AUTOTUNE = tf.data.experimental.AUTOTUNE

train_data = tf.data.Dataset.from_tensor_slices((x_train, y_train))

train_data = train_data.shuffle(5000).repeat(training_steps).batch(batch_size).prefetch(buffer_size=AUTOTUNE)

4.定义卷积层及池化层

@tf.function
def conv2d(x, W, b, strides=1):
    x = tf.nn.conv2d(x, W, strides=[1, strides, strides, 1], padding='SAME')
    x = tf.nn.bias_add(x, b)
    return tf.nn.relu(x)

@tf.function
def maxpool2d(x, k=2):
    return tf.nn.max_pool(x, ksize=[1, k, k, 1], strides=[1, k, k, 1], padding='SAME')

@tf.function

def conv2d(x, W, b, strides=1):

x = tf.nn.conv2d(x, W, strides=[1, strides, strides, 1], padding='SAME')

x = tf.nn.bias_add(x, b)

return tf.nn.relu(x)

@tf.function

def maxpool2d(x, k=2):

return tf.nn.max_pool(x, ksize=[1, k, k, 1], strides=[1, k, k, 1], padding='SAME')

5.构建模型

class CNNModel(tf.Module):
    def __init__(self,name = None):
        super(CNNModel, self).__init__(name=name)
        self.w1 = tf.Variable(random_normal([3, 3, 3, conv1_filters]))#[k_width, k_height, input_chn, output_chn]
        self.b1 = tf.Variable(tf.zeros([conv1_filters]))
        #输入通道：32，输出通道：64，卷积后图像尺寸不变，依然是16x16
        self.w2 = tf.Variable(random_normal([3, 3, conv1_filters, conv2_filters]))
        self.b2 = tf.Variable(tf.zeros([conv2_filters]))
        #将池第2个池化层的64个8x8的图像转换为一维的向量，长度是 64*8*8=4096
        self.w3 = tf.Variable(random_normal([4096, fc1_units]))
        self.b3 = tf.Variable(tf.zeros([fc1_units]))
        self.wout = tf.Variable(random_normal([fc1_units, num_classes]))
        self.bout = tf.Variable(tf.zeros([num_classes]))

     
    # 正向传播
    @tf.function
    def __call__(self,x):
        conv1 = conv2d(x, self.w1, self.b1)
        pool1 = maxpool2d(conv1, k=2)  #将32x32图像缩小为16x16，池化不改变通道数量，因此依然是32个
        conv2 = conv2d(pool1, self.w2, self.b2)
        pool2 = maxpool2d(conv2, k=2)
        flat = tf.reshape(pool2, [-1, self.w3.get_shape().as_list()[0]])
        fc1 = tf.add(tf.matmul(flat, self.w3), self.b3)
        fc1 = tf.nn.relu(fc1)
        out = tf.add(tf.matmul(fc1, self.wout), self.bout)
    
        return tf.nn.softmax(out) 
    
    # 损失函数(二元交叉熵)
    @tf.function
    def cross_entropy(self,y_pred, y_true):
        y_pred = tf.clip_by_value(y_pred, 1e-9, 1.)
        loss_ = tf.keras.losses.sparse_categorical_crossentropy(y_true=y_true, y_pred=y_pred)
    
        return tf.reduce_mean(loss_)
    
    # 评估指标(准确率)
    @tf.function
    def accuracy(self,y_pred, y_true):
        correct_prediction = tf.equal(tf.argmax(y_pred, 1), tf.reshape(tf.cast(y_true, tf.int64), [-1]))
        return tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
   
    
model = CNNModel()
#通过添加属性的方法进行封装
model.optimizer = tf.optimizers.Adam(learning_rate)

class CNNModel(tf.Module):

def __init__(self,name = None):

super(CNNModel, self).__init__(name=name)

self.w1 = tf.Variable(random_normal([3, 3, 3, conv1_filters]))#[k_width, k_height, input_chn, output_chn]

self.b1 = tf.Variable(tf.zeros([conv1_filters]))

#输入通道：32，输出通道：64，卷积后图像尺寸不变，依然是16x16

self.w2 = tf.Variable(random_normal([3, 3, conv1_filters, conv2_filters]))

self.b2 = tf.Variable(tf.zeros([conv2_filters]))

#将池第2个池化层的64个8x8的图像转换为一维的向量，长度是 64*8*8=4096

self.w3 = tf.Variable(random_normal([4096, fc1_units]))

self.b3 = tf.Variable(tf.zeros([fc1_units]))

self.wout = tf.Variable(random_normal([fc1_units, num_classes]))

self.bout = tf.Variable(tf.zeros([num_classes]))

# 正向传播

@tf.function

def __call__(self,x):

conv1 = conv2d(x, self.w1, self.b1)

pool1 = maxpool2d(conv1, k=2) #将32x32图像缩小为16x16，池化不改变通道数量，因此依然是32个

conv2 = conv2d(pool1, self.w2, self.b2)

pool2 = maxpool2d(conv2, k=2)

flat = tf.reshape(pool2, [-1, self.w3.get_shape().as_list()[0]])

fc1 = tf.add(tf.matmul(flat, self.w3), self.b3)

fc1 = tf.nn.relu(fc1)

out = tf.add(tf.matmul(fc1, self.wout), self.bout)

return tf.nn.softmax(out)

# 损失函数(二元交叉熵)

@tf.function

def cross_entropy(self,y_pred, y_true):

y_pred = tf.clip_by_value(y_pred, 1e-9, 1.)

loss_ = tf.keras.losses.sparse_categorical_crossentropy(y_true=y_true, y_pred=y_pred)

return tf.reduce_mean(loss_)

# 评估指标(准确率)

@tf.function

def accuracy(self,y_pred, y_true):

correct_prediction = tf.equal(tf.argmax(y_pred, 1), tf.reshape(tf.cast(y_true, tf.int64), [-1]))

return tf.reduce_mean(tf.cast(correct_prediction, tf.float32))

model = CNNModel()

#通过添加属性的方法进行封装

model.optimizer = tf.optimizers.Adam(learning_rate)

6.定义训练模型函数

自定训练过程：
（1）打开一个遍历各epoch的for循环
（2）对于每个epoch，打开一个分批遍历数据集的 for 循环
（3）对于每个批次，打开一个 GradientTape() 作用域
（4）在此作用域内，调用模型（前向传递）并计算损失
（5）在作用域之外，检索模型权重相对于损失的梯度
（6）根据梯度使用优化器来更新模型的权重
（7）评估模型指标

@tf.function
def train_step(model, features, labels):
    
    # 正向传播求损失
    with tf.GradientTape() as tape:
        predictions = model(features)
        loss = model.cross_entropy(predictions,labels) 
        
    # 反向传播求梯度
    grads = tape.gradient(loss, model.trainable_variables)
    
    # 执行梯度下降   
    model.optimizer.apply_gradients(zip(grads, model.trainable_variables))        
    # 计算评估指标
    metric = model.accuracy(predictions,labels)
    
    return loss, metric

train_loss_list1 = []
train_acc_list1 = []
def train_model(model,train_data,training_steps,display_step):
    #for epoch in tf.range(1,epochs+1): 
    for step, (batch_x, batch_y) in enumerate(train_data.take(training_steps), 1):
        starttime=time.time()
        loss,metric = train_step(model,batch_x,batch_y)
        
        if step % display_step == 0:
            #printbar()
            train_loss_list1.append(loss)
            train_acc_list1.append(metric)
            tf.print("step ={},loss = {:.4f},accuracy ={:.4f} ,times={:.4f}".format(step,loss,metric,(time.time() - starttime)))

@tf.function

def train_step(model, features, labels):

# 正向传播求损失

with tf.GradientTape() as tape:

predictions = model(features)

loss = model.cross_entropy(predictions,labels)

# 反向传播求梯度

grads = tape.gradient(loss, model.trainable_variables)

# 执行梯度下降

model.optimizer.apply_gradients(zip(grads, model.trainable_variables))

# 计算评估指标

metric = model.accuracy(predictions,labels)

return loss, metric

train_loss_list1 = []

train_acc_list1 = []

def train_model(model,train_data,training_steps,display_step):

#for epoch in tf.range(1,epochs+1):

for step, (batch_x, batch_y) in enumerate(train_data.take(training_steps), 1):

starttime=time.time()

loss,metric = train_step(model,batch_x,batch_y)

if step % display_step == 0:

#printbar()

train_loss_list1.append(loss)

train_acc_list1.append(metric)

tf.print("step ={},loss = {:.4f},accuracy ={:.4f} ,times={:.4f}".format(step,loss,metric,(time.time() - starttime)))

train_model(model,train_data,training_steps,display_step)
运行结果
step =1000,loss = 1.3011,accuracy =0.5781 ,times=0.0000
step =2000,loss = 1.2720,accuracy =0.6094 ,times=0.0000
step =3000,loss = 1.2153,accuracy =0.5469 ,times=0.0000
step =4000,loss = 0.8636,accuracy =0.7500 ,times=0.0000
step =5000,loss = 0.7936,accuracy =0.7500 ,times=0.0000
step =6000,loss = 0.9527,accuracy =0.6875 ,times=0.0156
step =7000,loss = 0.9352,accuracy =0.7344 ,times=0.0000
step =8000,loss = 0.6892,accuracy =0.7969 ,times=0.0000
step =9000,loss = 0.7949,accuracy =0.7031 ,times=0.0000
step =10000,loss = 0.4768,accuracy =0.8438 ,times=0.0000
step =11000,loss = 0.7983,accuracy =0.7188 ,times=0.0000
step =12000,loss = 0.5601,accuracy =0.8281 ,times=0.0000
step =13000,loss = 0.7934,accuracy =0.7031 ,times=0.0000
step =14000,loss = 0.6450,accuracy =0.8438 ,times=0.0000
step =15000,loss = 0.5681,accuracy =0.7656 ,times=0.0000
step =16000,loss = 0.5413,accuracy =0.8125 ,times=0.0000
step =17000,loss = 0.3914,accuracy =0.8438 ,times=0.0000
step =18000,loss = 0.3687,accuracy =0.8906 ,times=0.0000
step =19000,loss = 0.4534,accuracy =0.8750 ,times=0.0000
step =20000,loss = 0.3855,accuracy =0.8438 ,times=0.0000

从运行结果来看，运行时间快了很多！

7.可视化运行结果

plt.title('the train and validate')
plt.xlabel('Times')
plt.ylabel('Loss value')
plt.plot(train_loss_list, color=(1, 0, 0), label='Loss train')
plt.plot(train_acc_list, color=(0, 0, 1), label='Accuracy train')
plt.legend(loc='best')
plt.show()

plt.title('the train and validate')

plt.xlabel('Times')

plt.ylabel('Loss value')

plt.plot(train_loss_list, color=(1, 0, 0), label='Loss train')

plt.plot(train_acc_list, color=(0, 0, 1), label='Accuracy train')

plt.legend(loc='best')

plt.show()

8.测试模型

test_total_batch = int(len(x_test) / batch_size)
test_acc_sum = 0.0
for i in range(test_total_batch):
    test_image_batch = x_test[i*batch_size:(i+1)*batch_size]
    test_label_batch = y_test[i*batch_size:(i+1)*batch_size]
    pred = conv_net(test_image_batch)
    test_batch_acc = accuracy(pred,test_label_batch)
    test_acc_sum += test_batch_acc
test_acc = float(test_acc_sum / test_total_batch)
print("Test accuracy:{:.6f}".format(test_acc))

test_total_batch = int(len(x_test) / batch_size)

test_acc_sum = 0.0

for i in range(test_total_batch):

test_image_batch = x_test[i*batch_size:(i+1)*batch_size]

test_label_batch = y_test[i*batch_size:(i+1)*batch_size]

pred = conv_net(test_image_batch)

test_batch_acc = accuracy(pred,test_label_batch)

test_acc_sum += test_batch_acc

test_acc = float(test_acc_sum / test_total_batch)

print("Test accuracy:{:.6f}".format(test_acc))

运行结果
Test accuracy:0.720653
性能也有一定提升！

Tensorflow2构建模型-境界2

1.导入CIFAR-10数据集

1.1 导入需要的库

import os
import math
import numpy as np
import pickle as p
import tensorflow as tf
import matplotlib.pyplot as plt
%matplotlib inline

import os

import math

import numpy as np

import pickle as p

import tensorflow as tf

import matplotlib.pyplot as plt

%matplotlib inline

1.2 定义批量导入数据的函数

def load_CIFAR_batch(filename):
    """ load single batch of cifar """  
    with open(filename, 'rb')as f:
        # 一个样本由标签和图像数据组成
        #  (3072=32x32x3)
        # ...
        # 
        data_dict = p.load(f, encoding='bytes')
        images= data_dict[b'data']
        labels = data_dict[b'labels']
                
        # 把原始数据结构调整为: BCWH
        images = images.reshape(10000, 3, 32, 32)
        # tensorflow处理图像数据的结构：BWHC
        # 把通道数据C移动到最后一个维度
        images = images.transpose (0,2,3,1)
     
        labels = np.array(labels)
        
        return images, labels

def load_CIFAR_batch(filename):

""" load single batch of cifar """

with open(filename, 'rb')as f:

# 一个样本由标签和图像数据组成

# (3072=32x32x3)

# ...

data_dict = p.load(f, encoding='bytes')

images= data_dict[b'data']

labels = data_dict[b'labels']

# 把原始数据结构调整为: BCWH

images = images.reshape(10000, 3, 32, 32)

# tensorflow处理图像数据的结构：BWHC

# 把通道数据C移动到最后一个维度

images = images.transpose (0,2,3,1)

labels = np.array(labels)

return images, labels

1.3 定义加载数据函数

def load_CIFAR_data(data_dir):
    """load CIFAR data"""
 
    images_train=[]
    labels_train=[]
    for i in range(5):
        f=os.path.join(data_dir,'data_batch_%d' % (i+1))
        print('loading ',f)
        # 调用 load_CIFAR_batch( )获得批量的图像及其对应的标签
        image_batch,label_batch=load_CIFAR_batch(f)
        images_train.append(image_batch)
        labels_train.append(label_batch)
        Xtrain=np.concatenate(images_train)
        Ytrain=np.concatenate(labels_train)
        del image_batch ,label_batch
    
    Xtest,Ytest=load_CIFAR_batch(os.path.join(data_dir,'test_batch'))
    print('finished loadding CIFAR-10 data')
    
    # 返回训练集的图像和标签，测试集的图像和标签
    return (Xtrain,Ytrain),(Xtest,Ytest)

def load_CIFAR_data(data_dir):

"""load CIFAR data"""

images_train=[]

labels_train=[]

for i in range(5):

f=os.path.join(data_dir,'data_batch_%d' % (i+1))

print('loading ',f)

# 调用 load_CIFAR_batch( )获得批量的图像及其对应的标签

image_batch,label_batch=load_CIFAR_batch(f)

images_train.append(image_batch)

labels_train.append(label_batch)

Xtrain=np.concatenate(images_train)

Ytrain=np.concatenate(labels_train)

del image_batch ,label_batch

Xtest,Ytest=load_CIFAR_batch(os.path.join(data_dir,'test_batch'))

print('finished loadding CIFAR-10 data')

# 返回训练集的图像和标签，测试集的图像和标签

return (Xtrain,Ytrain),(Xtest,Ytest)

1.4 加载数据

data_dir = r'data\cifar-10-batches-py'
(x_train,y_train),(x_test,y_test) = load_CIFAR_data(data_dir)

1 2	data_dir = r'data\cifar-10-batches-py' (x_train,y_train),(x_test,y_test) = load_CIFAR_data(data_dir)

1.5 可视化加载数据

label_dict = {0:"airplane", 1:"
automobile", 2:"bird", 3:"cat", 4:"deer",
              5:"dog", 6:"frog", 7:"horse", 8:"ship", 9:"truck"}


def plot_images_labels(images, labels, num):
    total = len(images)
    fig = plt.gcf()
    fig.set_size_inches(15, math.ceil(num / 10) * 7)
    for i in range(0, num):
        choose_n = np.random.randint(0, total)
        
        ax = plt.subplot(math.ceil(num / 5), 5, 1 + i)
        ax.imshow(images[choose_n], cmap='binary')
        title = label_dict[labels[choose_n]]
        ax.set_title(title, fontsize=10)        
    plt.show()
	
plot_images_labels(x_train, y_train, 10)

label_dict = {0:"airplane", 1:"

automobile", 2:"bird", 3:"cat", 4:"deer",

5:"dog", 6:"frog", 7:"horse", 8:"ship", 9:"truck"}

def plot_images_labels(images, labels, num):

total = len(images)

fig = plt.gcf()

fig.set_size_inches(15, math.ceil(num / 10) * 7)

for i in range(0, num):

choose_n = np.random.randint(0, total)

ax = plt.subplot(math.ceil(num / 5), 5, 1 + i)

ax.imshow(images[choose_n], cmap='binary')

title = label_dict[labels[choose_n]]

ax.set_title(title, fontsize=10)

plt.show()

plot_images_labels(x_train, y_train, 10)

运行结果

2 .数据预处理并设置超参数

x_train = x_train.astype('float32') / 255.0
x_test = x_test.astype('float32') / 255.0

train_num = len(x_train)
num_classes = 10

learning_rate = 0.0002
batch_size = 64
training_steps = 20000
display_step = 1000

conv1_filters = 32
conv2_filters = 64
fc1_units = 256

x_train = x_train.astype('float32') / 255.0

x_test = x_test.astype('float32') / 255.0

train_num = len(x_train)

num_classes = 10

learning_rate = 0.0002

batch_size = 64

training_steps = 20000

display_step = 1000

conv1_filters = 32

conv2_filters = 64

fc1_units = 256

3.使用tf.data构建数据管道

AUTOTUNE = tf.data.experimental.AUTOTUNE
train_data = tf.data.Dataset.from_tensor_slices((x_train, y_train))
train_data = train_data.shuffle(5000).repeat(training_steps).batch(batch_size).prefetch(buffer_size=AUTOTUNE)

AUTOTUNE = tf.data.experimental.AUTOTUNE

train_data = tf.data.Dataset.from_tensor_slices((x_train, y_train))

train_data = train_data.shuffle(5000).repeat(training_steps).batch(batch_size).prefetch(buffer_size=AUTOTUNE)

4.定义卷积层及池化层

def conv2d(x, W, b, strides=1):
    #tf.nn.conv2d(input, filter, strides, padding, use_cudnn_on_gpu=None, name=None)
    x = tf.nn.conv2d(x, W, strides=[1, strides, strides, 1], padding='SAME')
    x = tf.nn.bias_add(x, b)
    return tf.nn.relu(x)


def maxpool2d(x, k=2):
    #tf.nn.max_pool(value, ksize, strides, padding, name=None) 
return tf.nn.max_pool(x, ksize=[1, k, k, 1], strides=[1, k, k, 1], padding='SAME')

def conv2d(x, W, b, strides=1):

#tf.nn.conv2d(input, filter, strides, padding, use_cudnn_on_gpu=None, name=None)

x = tf.nn.conv2d(x, W, strides=[1, strides, strides, 1], padding='SAME')

x = tf.nn.bias_add(x, b)

return tf.nn.relu(x)

def maxpool2d(x, k=2):

#tf.nn.max_pool(value, ksize, strides, padding, name=None)

return tf.nn.max_pool(x, ksize=[1, k, k, 1], strides=[1, k, k, 1], padding='SAME')

5.构建模型

class CNNModel(tf.Module):
    def __init__(self,name = None):
        super(CNNModel, self).__init__(name=name)
        self.w1 = tf.Variable(random_normal([3, 3, 3, conv1_filters]))#[k_width, k_height, input_chn, output_chn]
        self.b1 = tf.Variable(tf.zeros([conv1_filters]))
        #输入通道：32，输出通道：64，卷积后图像尺寸不变，依然是16x16
        self.w2 = tf.Variable(random_normal([3, 3, conv1_filters, conv2_filters]))
        self.b2 = tf.Variable(tf.zeros([conv2_filters]))
        #将池第2个池化层的64个8x8的图像转换为一维的向量，长度是 64*8*8=4096
        self.w3 = tf.Variable(random_normal([4096, fc1_units]))
        self.b3 = tf.Variable(tf.zeros([fc1_units]))
        self.wout = tf.Variable(random_normal([fc1_units, num_classes]))
        self.bout = tf.Variable(tf.zeros([num_classes]))

     
    # 正向传播    
    def __call__(self,x):
        conv1 = conv2d(x, self.w1, self.b1)
        pool1 = maxpool2d(conv1, k=2)  #将32x32图像缩小为16x16，池化不改变通道数量，因此依然是32个
        conv2 = conv2d(pool1, self.w2, self.b2)
        pool2 = maxpool2d(conv2, k=2)
        flat = tf.reshape(pool2, [-1, self.w3.get_shape().as_list()[0]])
        fc1 = tf.add(tf.matmul(flat, self.w3), self.b3)
        fc1 = tf.nn.relu(fc1)
        out = tf.add(tf.matmul(fc1, self.wout), self.bout)
    
        return tf.nn.softmax(out) 
    
    # 损失函数(二元交叉熵)
    #@tf.function(input_signature=[tf.TensorSpec(shape = [None,1], dtype = tf.float32),tf.TensorSpec(shape = [None,1], dtype = tf.float32)])  
    def cross_entropy(self,y_pred, y_true):
        y_pred = tf.clip_by_value(y_pred, 1e-9, 1.)
        loss_ = tf.keras.losses.sparse_categorical_crossentropy(y_true=y_true, y_pred=y_pred)
    
        return tf.reduce_mean(loss_)
    
    # 评估指标(准确率)    
    def accuracy(self,y_pred, y_true):
        correct_prediction = tf.equal(tf.argmax(y_pred, 1), tf.reshape(tf.cast(y_true, tf.int64), [-1]))
        return tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
   
    
model = CNNModel()
model.optimizer = tf.optimizers.Adam(learning_rate)

class CNNModel(tf.Module):

def __init__(self,name = None):

super(CNNModel, self).__init__(name=name)

self.w1 = tf.Variable(random_normal([3, 3, 3, conv1_filters]))#[k_width, k_height, input_chn, output_chn]

self.b1 = tf.Variable(tf.zeros([conv1_filters]))

#输入通道：32，输出通道：64，卷积后图像尺寸不变，依然是16x16

self.w2 = tf.Variable(random_normal([3, 3, conv1_filters, conv2_filters]))

self.b2 = tf.Variable(tf.zeros([conv2_filters]))

#将池第2个池化层的64个8x8的图像转换为一维的向量，长度是 64*8*8=4096

self.w3 = tf.Variable(random_normal([4096, fc1_units]))

self.b3 = tf.Variable(tf.zeros([fc1_units]))

self.wout = tf.Variable(random_normal([fc1_units, num_classes]))

self.bout = tf.Variable(tf.zeros([num_classes]))

# 正向传播

def __call__(self,x):

conv1 = conv2d(x, self.w1, self.b1)

pool1 = maxpool2d(conv1, k=2) #将32x32图像缩小为16x16，池化不改变通道数量，因此依然是32个

conv2 = conv2d(pool1, self.w2, self.b2)

pool2 = maxpool2d(conv2, k=2)

flat = tf.reshape(pool2, [-1, self.w3.get_shape().as_list()[0]])

fc1 = tf.add(tf.matmul(flat, self.w3), self.b3)

fc1 = tf.nn.relu(fc1)

out = tf.add(tf.matmul(fc1, self.wout), self.bout)

return tf.nn.softmax(out)

# 损失函数(二元交叉熵)

#@tf.function(input_signature=[tf.TensorSpec(shape = [None,1], dtype = tf.float32),tf.TensorSpec(shape = [None,1], dtype = tf.float32)])

def cross_entropy(self,y_pred, y_true):

y_pred = tf.clip_by_value(y_pred, 1e-9, 1.)

loss_ = tf.keras.losses.sparse_categorical_crossentropy(y_true=y_true, y_pred=y_pred)

return tf.reduce_mean(loss_)

# 评估指标(准确率)

def accuracy(self,y_pred, y_true):

correct_prediction = tf.equal(tf.argmax(y_pred, 1), tf.reshape(tf.cast(y_true, tf.int64), [-1]))

return tf.reduce_mean(tf.cast(correct_prediction, tf.float32))

model = CNNModel()

model.optimizer = tf.optimizers.Adam(learning_rate)

6.定义训练模型函数

自定义训练过程：
（1）打开一个遍历各epoch的for循环
（2）对于每个epoch，打开一个分批遍历数据集的 for 循环
（3）对于每个批次，打开一个 GradientTape() 作用域
（4）在此作用域内，调用模型（前向传递）并计算损失
（5）在作用域之外，检索模型权重相对于损失的梯度
（6）根据梯度使用优化器来更新模型的权重
（7）评估模型指标

def train_step(model, features, labels):
    
    # 正向传播求损失
    with tf.GradientTape() as tape:
        predictions = model(features)
        loss = model.cross_entropy(predictions,labels) 
        
    # 反向传播求梯度
    grads = tape.gradient(loss, model.trainable_variables)
    
    # 执行梯度下降   
    model.optimizer.apply_gradients(zip(grads, model.trainable_variables))        
    # 计算评估指标
    metric = model.accuracy(predictions,labels)
    
    return loss, metric

train_loss_list1 = []
train_acc_list1 = []
def train_model(model,train_data,training_steps,display_step):
    #for epoch in tf.range(1,epochs+1): 
    for step, (batch_x, batch_y) in enumerate(train_data.take(training_steps), 1):
        starttime=time.time()
        loss,metric = train_step(model,batch_x,batch_y)
        
        if step % display_step == 0:
            #printbar()
            train_loss_list1.append(loss)
            train_acc_list1.append(metric)
            tf.print("step ={},loss = {:.4f},accuracy ={:.4f} ,times={:.4f}".format(step,loss,metric,(time.time() - starttime)))

def train_step(model, features, labels):

# 正向传播求损失

with tf.GradientTape() as tape:

predictions = model(features)

loss = model.cross_entropy(predictions,labels)

# 反向传播求梯度

grads = tape.gradient(loss, model.trainable_variables)

# 执行梯度下降

model.optimizer.apply_gradients(zip(grads, model.trainable_variables))

# 计算评估指标

metric = model.accuracy(predictions,labels)

return loss, metric

train_loss_list1 = []

train_acc_list1 = []

def train_model(model,train_data,training_steps,display_step):

#for epoch in tf.range(1,epochs+1):

for step, (batch_x, batch_y) in enumerate(train_data.take(training_steps), 1):

starttime=time.time()

loss,metric = train_step(model,batch_x,batch_y)

if step % display_step == 0:

#printbar()

train_loss_list1.append(loss)

train_acc_list1.append(metric)

tf.print("step ={},loss = {:.4f},accuracy ={:.4f} ,times={:.4f}".format(step,loss,metric,(time.time() - starttime)))

train_model(model,train_data,training_steps,display_step)
运行结果
step =1000,loss = 1.2807,accuracy =0.5781 ,times=0.0080
step =2000,loss = 1.1832,accuracy =0.6562 ,times=0.0080
step =3000,loss = 0.9727,accuracy =0.6562 ,times=0.0080
step =4000,loss = 1.0398,accuracy =0.6406 ,times=0.0050
step =5000,loss = 0.8615,accuracy =0.6406 ,times=0.0156
step =6000,loss = 0.7207,accuracy =0.7188 ,times=0.0000
step =7000,loss = 1.0945,accuracy =0.5938 ,times=0.0090
step =8000,loss = 0.7337,accuracy =0.7656 ,times=0.0080
step =9000,loss = 0.5792,accuracy =0.7812 ,times=0.0080
step =10000,loss = 0.7154,accuracy =0.7500 ,times=0.0080
step =11000,loss = 0.6398,accuracy =0.8125 ,times=0.0156
step =12000,loss = 0.6413,accuracy =0.7500 ,times=0.0080
step =13000,loss = 0.5555,accuracy =0.7812 ,times=0.0156
step =14000,loss = 0.6729,accuracy =0.8281 ,times=0.0156
step =15000,loss = 0.5163,accuracy =0.7500 ,times=0.0156
step =16000,loss = 0.5521,accuracy =0.7969 ,times=0.0000
step =17000,loss = 0.4475,accuracy =0.8594 ,times=0.0156
step =18000,loss = 0.3158,accuracy =0.8594 ,times=0.0080
step =19000,loss = 0.3829,accuracy =0.9062 ,times=0.0090
step =20000,loss = 0.2731,accuracy =0.9062 ,times=0.0080

7.可视化运行结果

plt.title('the train and validate')
plt.xlabel('Times')
plt.ylabel('Loss value')
plt.plot(train_loss_list, color=(1, 0, 0), label='Loss train')
plt.plot(train_acc_list, color=(0, 0, 1), label='Accuracy train')
plt.legend(loc='best')
plt.show()

plt.title('the train and validate')

plt.xlabel('Times')

plt.ylabel('Loss value')

plt.plot(train_loss_list, color=(1, 0, 0), label='Loss train')

plt.plot(train_acc_list, color=(0, 0, 1), label='Accuracy train')

plt.legend(loc='best')

plt.show()

8.测试型

test_total_batch = int(len(x_test) / batch_size)
test_acc_sum = 0.0
for i in range(test_total_batch):
    test_image_batch = x_test[i*batch_size:(i+1)*batch_size]
    test_label_batch = y_test[i*batch_size:(i+1)*batch_size]
    pred = conv_net(test_image_batch)
    test_batch_acc = accuracy(pred,test_label_batch)
    test_acc_sum += test_batch_acc
test_acc = float(test_acc_sum / test_total_batch)
print("Test accuracy:{:.6f}".format(test_acc))

test_total_batch = int(len(x_test) / batch_size)

test_acc_sum = 0.0

for i in range(test_total_batch):

test_image_batch = x_test[i*batch_size:(i+1)*batch_size]

test_label_batch = y_test[i*batch_size:(i+1)*batch_size]

pred = conv_net(test_image_batch)

test_batch_acc = accuracy(pred,test_label_batch)

test_acc_sum += test_batch_acc

test_acc = float(test_acc_sum / test_total_batch)

print("Test accuracy:{:.6f}".format(test_acc))

运行结果
Test accuracy:0.712640

TensorFlow2构建模型-境界1

1.导入CIFAR-10数据集

1.1 导入需要的库

import os
import math
import numpy as np
import pickle as p
import tensorflow as tf
import matplotlib.pyplot as plt
%matplotlib inline

import os

import math

import numpy as np

import pickle as p

import tensorflow as tf

import matplotlib.pyplot as plt

%matplotlib inline

1.2 定义批量导入数据的函数

def load_CIFAR_batch(filename):
    """ load single batch of cifar """  
    with open(filename, 'rb')as f:
        # 一个样本由标签和图像数据组成
        #  (3072=32x32x3)
        # ...
        # 
        data_dict = p.load(f, encoding='bytes')
        images= data_dict[b'data']
        labels = data_dict[b'labels']
                
        # 把原始数据结构调整为: BCWH
        images = images.reshape(10000, 3, 32, 32)
        # tensorflow处理图像数据的结构：BWHC
        # 把通道数据C移动到最后一个维度
        images = images.transpose (0,2,3,1)
     
        labels = np.array(labels)
        
        return images, labels

def load_CIFAR_batch(filename):

""" load single batch of cifar """

with open(filename, 'rb')as f:

# 一个样本由标签和图像数据组成

# (3072=32x32x3)

# ...

data_dict = p.load(f, encoding='bytes')

images= data_dict[b'data']

labels = data_dict[b'labels']

# 把原始数据结构调整为: BCWH

images = images.reshape(10000, 3, 32, 32)

# tensorflow处理图像数据的结构：BWHC

# 把通道数据C移动到最后一个维度

images = images.transpose (0,2,3,1)

labels = np.array(labels)

return images, labels

1.3 定义加载数据函数

def load_CIFAR_data(data_dir):
    """load CIFAR data"""
 
    images_train=[]
    labels_train=[]
    for i in range(5):
        f=os.path.join(data_dir,'data_batch_%d' % (i+1))
        print('loading ',f)
        # 调用 load_CIFAR_batch( )获得批量的图像及其对应的标签
        image_batch,label_batch=load_CIFAR_batch(f)
        images_train.append(image_batch)
        labels_train.append(label_batch)
        Xtrain=np.concatenate(images_train)
        Ytrain=np.concatenate(labels_train)
        del image_batch ,label_batch
    
    Xtest,Ytest=load_CIFAR_batch(os.path.join(data_dir,'test_batch'))
    print('finished loadding CIFAR-10 data')
    
    # 返回训练集的图像和标签，测试集的图像和标签
    return (Xtrain,Ytrain),(Xtest,Ytest)

def load_CIFAR_data(data_dir):

"""load CIFAR data"""

images_train=[]

labels_train=[]

for i in range(5):

f=os.path.join(data_dir,'data_batch_%d' % (i+1))

print('loading ',f)

# 调用 load_CIFAR_batch( )获得批量的图像及其对应的标签

image_batch,label_batch=load_CIFAR_batch(f)

images_train.append(image_batch)

labels_train.append(label_batch)

Xtrain=np.concatenate(images_train)

Ytrain=np.concatenate(labels_train)

del image_batch ,label_batch

Xtest,Ytest=load_CIFAR_batch(os.path.join(data_dir,'test_batch'))

print('finished loadding CIFAR-10 data')

# 返回训练集的图像和标签，测试集的图像和标签

return (Xtrain,Ytrain),(Xtest,Ytest)

1.4 加载数据

data_dir = r'data\cifar-10-batches-py'
(x_train,y_train),(x_test,y_test) = load_CIFAR_data(data_dir)

1 2	data_dir = r'data\cifar-10-batches-py' (x_train,y_train),(x_test,y_test) = load_CIFAR_data(data_dir)

1.5 可视化加载数据

label_dict = {0:"airplane", 1:"
automobile", 2:"bird", 3:"cat", 4:"deer",
              5:"dog", 6:"frog", 7:"horse", 8:"ship", 9:"truck"}


def plot_images_labels(images, labels, num):
    total = len(images)
    fig = plt.gcf()
    fig.set_size_inches(15, math.ceil(num / 10) * 7)
    for i in range(0, num):
        choose_n = np.random.randint(0, total)
        
        ax = plt.subplot(math.ceil(num / 5), 5, 1 + i)
        ax.imshow(images[choose_n], cmap='binary')
        title = label_dict[labels[choose_n]]
        ax.set_title(title, fontsize=10)        
    plt.show()
	
plot_images_labels(x_train, y_train, 10)

label_dict = {0:"airplane", 1:"

automobile", 2:"bird", 3:"cat", 4:"deer",

5:"dog", 6:"frog", 7:"horse", 8:"ship", 9:"truck"}

def plot_images_labels(images, labels, num):

total = len(images)

fig = plt.gcf()

fig.set_size_inches(15, math.ceil(num / 10) * 7)

for i in range(0, num):

choose_n = np.random.randint(0, total)

ax = plt.subplot(math.ceil(num / 5), 5, 1 + i)

ax.imshow(images[choose_n], cmap='binary')

title = label_dict[labels[choose_n]]

ax.set_title(title, fontsize=10)

plt.show()

plot_images_labels(x_train, y_train, 10)

运行结果

2 .数据预处理并设置超参数

x_train = x_train.astype('float32') / 255.0
x_test = x_test.astype('float32') / 255.0

train_num = len(x_train)
num_classes = 10

learning_rate = 0.0002
batch_size = 64
training_steps = 20000
display_step = 1000

conv1_filters = 32
conv2_filters = 64
fc1_units = 256

x_train = x_train.astype('float32') / 255.0

x_test = x_test.astype('float32') / 255.0

train_num = len(x_train)

num_classes = 10

learning_rate = 0.0002

batch_size = 64

training_steps = 20000

display_step = 1000

conv1_filters = 32

conv2_filters = 64

fc1_units = 256

3.使用tf,data构建数据管道

AUTOTUNE = tf.data.experimental.AUTOTUNE
train_data = tf.data.Dataset.from_tensor_slices((x_train, y_train))
train_data = train_data.shuffle(5000).repeat(training_steps).batch(batch_size).prefetch(buffer_size=AUTOTUNE)

AUTOTUNE = tf.data.experimental.AUTOTUNE

train_data = tf.data.Dataset.from_tensor_slices((x_train, y_train))

train_data = train_data.shuffle(5000).repeat(training_steps).batch(batch_size).prefetch(buffer_size=AUTOTUNE)

4.定义卷积层及池化层

def conv2d(x, W, b, strides=1):
    #tf.nn.conv2d(input, filter, strides, padding, use_cudnn_on_gpu=None, name=None)
    x = tf.nn.conv2d(x, W, strides=[1, strides, strides, 1], padding='SAME')
    x = tf.nn.bias_add(x, b)
    return tf.nn.relu(x)


def maxpool2d(x, k=2):
    #tf.nn.max_pool(value, ksize, strides, padding, name=None) 
return tf.nn.max_pool(x, ksize=[1, k, k, 1], strides=[1, k, k, 1], padding='SAME')

def conv2d(x, W, b, strides=1):

#tf.nn.conv2d(input, filter, strides, padding, use_cudnn_on_gpu=None, name=None)

x = tf.nn.conv2d(x, W, strides=[1, strides, strides, 1], padding='SAME')

x = tf.nn.bias_add(x, b)

return tf.nn.relu(x)

def maxpool2d(x, k=2):

#tf.nn.max_pool(value, ksize, strides, padding, name=None)

return tf.nn.max_pool(x, ksize=[1, k, k, 1], strides=[1, k, k, 1], padding='SAME')

5.定义并初始化权重参数

random_normal = tf.initializers.RandomNormal()

weights = { 
    'wc1': tf.Variable(random_normal([3, 3, 3, conv1_filters])),#[k_width, k_height, input_chn, output_chn]
    'wc2': tf.Variable(random_normal([3, 3, conv1_filters, conv2_filters])),#输入通道：32，输出通道：64，卷积后图像尺寸不变，依然是16x16
    'wd1': tf.Variable(random_normal([4096, fc1_units])),#将池第2个池化层的64个8x8的图像转换为一维的向量，长度是 64*8*8=4096
    'out': tf.Variable(random_normal([fc1_units, num_classes]))
}

biases = {
    'bc1': tf.Variable(tf.zeros([conv1_filters])),
    'bc2': tf.Variable(tf.zeros([conv2_filters])),
    'bd1': tf.Variable(tf.zeros([fc1_units])),
    'out': tf.Variable(tf.zeros([num_classes]))
}

random_normal = tf.initializers.RandomNormal()

weights = {

'wc1': tf.Variable(random_normal([3, 3, 3, conv1_filters])),#[k_width, k_height, input_chn, output_chn]

'wc2': tf.Variable(random_normal([3, 3, conv1_filters, conv2_filters])),#输入通道：32，输出通道：64，卷积后图像尺寸不变，依然是16x16

'wd1': tf.Variable(random_normal([4096, fc1_units])),#将池第2个池化层的64个8x8的图像转换为一维的向量，长度是 64*8*8=4096

'out': tf.Variable(random_normal([fc1_units, num_classes]))

}

biases = {

'bc1': tf.Variable(tf.zeros([conv1_filters])),

'bc2': tf.Variable(tf.zeros([conv2_filters])),

'bd1': tf.Variable(tf.zeros([fc1_units])),

'out': tf.Variable(tf.zeros([num_classes]))

}

6.构建模型

def conv_net(x):
    x = tf.reshape(x, [-1, 32, 32, 3])## 32x32图像，通道为3（RGB）
    conv1 = conv2d(x, weights['wc1'], biases['bc1'])
    pool1 = maxpool2d(conv1, k=2)  #将32x32图像缩小为16x16，池化不改变通道数量，因此依然是32个
    conv2 = conv2d(pool1, weights['wc2'], biases['bc2'])
    pool2 = maxpool2d(conv2, k=2)
    flat = tf.reshape(pool2, [-1, weights['wd1'].get_shape().as_list()[0]])
    fc1 = tf.add(tf.matmul(flat, weights['wd1']), biases['bd1'])
    fc1 = tf.nn.relu(fc1)
    out = tf.add(tf.matmul(fc1, weights['out']), biases['out'])
    
    return tf.nn.softmax(out)

def conv_net(x):

x = tf.reshape(x, [-1, 32, 32, 3])## 32x32图像，通道为3（RGB）

conv1 = conv2d(x, weights['wc1'], biases['bc1'])

pool1 = maxpool2d(conv1, k=2) #将32x32图像缩小为16x16，池化不改变通道数量，因此依然是32个

conv2 = conv2d(pool1, weights['wc2'], biases['bc2'])

pool2 = maxpool2d(conv2, k=2)

flat = tf.reshape(pool2, [-1, weights['wd1'].get_shape().as_list()[0]])

fc1 = tf.add(tf.matmul(flat, weights['wd1']), biases['bd1'])

fc1 = tf.nn.relu(fc1)

out = tf.add(tf.matmul(fc1, weights['out']), biases['out'])

return tf.nn.softmax(out)

7.定义损失函数、评估函数

def cross_entropy(y_pred, y_true):
    y_pred = tf.clip_by_value(y_pred, 1e-9, 1.)
    loss_ = tf.keras.losses.sparse_categorical_crossentropy(y_true=y_true, y_pred=y_pred)
    
    return tf.reduce_mean(loss_)


def accuracy(y_pred, y_true):
    correct_prediction = tf.equal(tf.argmax(y_pred, 1), tf.reshape(tf.cast(y_true, tf.int64), [-1]))

    return tf.reduce_mean(tf.cast(correct_prediction, tf.float32))

optimizer = tf.optimizers.Adam(learning_rate)

def cross_entropy(y_pred, y_true):

y_pred = tf.clip_by_value(y_pred, 1e-9, 1.)

loss_ = tf.keras.losses.sparse_categorical_crossentropy(y_true=y_true, y_pred=y_pred)

return tf.reduce_mean(loss_)

def accuracy(y_pred, y_true):

correct_prediction = tf.equal(tf.argmax(y_pred, 1), tf.reshape(tf.cast(y_true, tf.int64), [-1]))

return tf.reduce_mean(tf.cast(correct_prediction, tf.float32))

optimizer = tf.optimizers.Adam(learning_rate)

8.定义梯度计算函数

自定义梯度计算过程：
（1）打开一个 GradientTape() 作用域
（2）在此作用域内，调用模型（正向传播）并计算损失
（3）在作用域之外，检索模型权重相对于损失的梯度
（4）根据梯度使用优化器来更新模型的权重
（5）利用优化器进行反向传播（更新梯度）

[{"metadata":{"trusted":false},"cell_type":"code","source":"def run_optimization(x, y):\n    with tf.GradientTape() as g:\n        pred = conv_net(x)\n        loss = cross_entropy(pred, y)\n        \n    trainable_variables = list(weights.values()) + list(biases.values())\n    gradients = g.gradient(loss, trainable_variables)\n    optimizer.apply_gradients(zip(gradients, trainable_variables))\n","execution_count":13,"outputs":[]}]

[{"metadata":{"trusted":false},"cell_type":"code","source":"def run_optimization(x, y):\n with tf.GradientTape() as g:\n pred = conv_net(x)\n loss = cross_entropy(pred, y)\n \n trainable_variables = list(weights.values()) + list(biases.values())\n gradients = g.gradient(loss, trainable_variables)\n optimizer.apply_gradients(zip(gradients, trainable_variables))\n","execution_count":13,"outputs":[]}]

9.训练模型

train_loss_list = []
train_acc_list = []

for step, (batch_x, batch_y) in enumerate(train_data.take(training_steps), 1):
    run_optimization(batch_x, batch_y)
    
    if step % display_step == 0:
        pred = conv_net(batch_x)
        loss = cross_entropy(pred, batch_y)
        acc = accuracy(pred, batch_y)
        train_loss_list.append(loss)
        train_acc_list.append(acc)
        print("step: %i, loss: %f, accuracy: %f" % (step, loss, acc))

train_loss_list = []

train_acc_list = []

for step, (batch_x, batch_y) in enumerate(train_data.take(training_steps), 1):

run_optimization(batch_x, batch_y)

if step % display_step == 0:

pred = conv_net(batch_x)

loss = cross_entropy(pred, batch_y)

acc = accuracy(pred, batch_y)

train_loss_list.append(loss)

train_acc_list.append(acc)

print("step: %i, loss: %f, accuracy: %f" % (step, loss, acc))

10.可视化运行结果

plt.title('the train and validate')
plt.xlabel('Times')
plt.ylabel('Loss value')
plt.plot(train_loss_list, color=(1, 0, 0), label='Loss train')
plt.plot(train_acc_list, color=(0, 0, 1), label='Accuracy train')
plt.legend(loc='best')
plt.show()

plt.title('the train and validate')

plt.xlabel('Times')

plt.ylabel('Loss value')

plt.plot(train_loss_list, color=(1, 0, 0), label='Loss train')

plt.plot(train_acc_list, color=(0, 0, 1), label='Accuracy train')

plt.legend(loc='best')

plt.show()

运行结果

11.测试模型

test_total_batch = int(len(x_test) / batch_size)
test_acc_sum = 0.0
for i in range(test_total_batch):
    test_image_batch = x_test[i*batch_size:(i+1)*batch_size]
    test_label_batch = y_test[i*batch_size:(i+1)*batch_size]
    pred = conv_net(test_image_batch)
    test_batch_acc = accuracy(pred,test_label_batch)
    test_acc_sum += test_batch_acc
test_acc = float(test_acc_sum / test_total_batch)
print("Test accuracy:{:.6f}".format(test_acc))

test_total_batch = int(len(x_test) / batch_size)

test_acc_sum = 0.0

for i in range(test_total_batch):

test_image_batch = x_test[i*batch_size:(i+1)*batch_size]

test_label_batch = y_test[i*batch_size:(i+1)*batch_size]

pred = conv_net(test_image_batch)

test_batch_acc = accuracy(pred,test_label_batch)

test_acc_sum += test_batch_acc

test_acc = float(test_acc_sum / test_total_batch)

print("Test accuracy:{:.6f}".format(test_acc))

运行结果
Test accuracy:0.704327

TensorFlow2构建模型的五个境界

一、纯手工

定制化程度高，非常灵活，如自定义层、损失函数、优化器、性能评估、训练模型方法等，但编程难度较高。
下载所有实例使用数据（提取码为：fg29）

具体实例

二、继承tf.Module

定制化程度高，如自定义层、损失函数、优化器、性能评估、训练模型方法等，且可以使用保存或恢复模型。
具体实例

三、继承tf.Module且使用AutoGraph

定制化程度高，如自定层、损失函数、优化器、性能评估、训练模型方法等，且可以使用保存或恢复模型。通过使用Autograph可达到静态图的性能，极大提升训练模型效率。不过使用Autograph时要注意对应的编码规范，如被@tf,function装饰的函数尽可能tensorflow函数，如tf.print,tf.range等（控制语句无此限制），避免在被@tf,function装饰的函数内部定义tf.Variabe等。
具体实例

四、网络层和模型使用子类方法构建

用子类方法构建网络层、网络模型是tensorflow2才有的，这种风格类似于PyTorch构建方法。TensorFlow1构建网络一般使用Sequential按层顺序构建模型，使用函数式API构建任意结构模型。当然，这两种方法在TensorFlow2也是常用方法。构建网络层时继承
tf.keras.layers.Layer，构建网络模型继承tf.keras.Model，然后使用compile编辑模型，用fit训练模型。这样构建的网络既有一定的灵活性，又不失简洁性。
具体实例

五、使用tf,keras的Sequential模型按层顺序构建模型

最简洁，构建模型很简单，可是直接使用内置的层、损失函数、优化器、评估指标等，训练也可直接使用函数fit即可。Keras 提供了称为tf.keras.Model 的全功能模型类。它继承自 tf.keras.layers.Layer，使用SaveModel保存模型，可以非常方便移植到其他非Python环境（具体环境可参考下图）。构建这个模型相对较简单，有兴趣的朋友作为练习。

构建机器学习的一般步骤（基于TensorFlow2）

如何有效使用tf.data(基于TensorFlow2)？

1.机器学习一般流程

机器学习、深度学习的一般流程：
（1）分析业务需求
（2）确定数据源
（3）构建数据处理管道（pipeline）
（4）构建模型
（5）训练模型
（6）评估、优化模型
其中构建数据处理管道，在整个过程中，从时间上来说往往占60%左右，面对大数据其挑战更大。如何解决质量问题？如何解决内存瓶颈问题？如何解决处理效率问题等等。
接下来我们重点介绍如何使用TensorFlow2提供的tf.data工具有效构建数据流。

2.为何要构建数据管道？

使用tf.data API，我们可以使用简单的代码来构建复杂的输入 pipeline。
实现从从内存读取数据、从分布式文件系统中读取数据、进行预处理、合成为 batch、训练中使用数据集等。使用 tf.data API 可以轻松处理大量数据、不同的数据格式以及复杂的转换。
如果数据比较小，我们可以一次性处理后直接加入内存就可以了；但如果数据比较大，而且在数据训练过程中还需要一些动态的处理方法，如分批处理、添加数据增强方法、数据采样等等，此时，通过构建一个数据流就显得非常必要。
数据流可有效提高我们管理数据得效率，此外还可以帮助我们解耦数据的预处理和数据执行的过程，能够帮我们更高效的应用硬件资源，例如当分布式训练的时候，一个好的数据流能够帮我们高效的分发数据到不同的硬件上，从而提高整体的训练效率。
一个合理的数据流，能够让你模型训练更加的高效。数据流的本质就是 ETL。一般来说，数据流由三部分组成，具体如下：
（1）抽取、初始化源数据 (E:即Extract)
（2）添加各种预处理过程 (T: 即Transform)
（3）遍历数据流，把大数据导入模型、训练数据等 (L:Load)
在 Tensorflow2里，我们使用 tf.data 来构建数据管道。

3.tf.data简介

tf.data是TensorFlow提供的构建数据管道的一个工具，与PyTorch的utils.data类似，使用tf.data构建数据集（Dataset），构建和管理数据管道非常方便，它提供了很多操作，如：
shuffle、repeat、map、batch、padded_batch、prefetch等等，这些操作功能很实用，但使用的顺序是有讲究的，如果次序不当，将严重影响数据流的效率和质量，这些操作的顺利大致为：
（1）map->shuffle -> repeat -> map(parse) -> batch -> prefetch ；
（2）有些 map 操作放在 batch 前，有些 map 操作放在 batch 后；
（3）尽量把进行数据过滤和采样放数据流的前面，以提高后续处理效率；
（4）使用 AUTOTUNE 来设置并行执行的数量，不要去手动调节；
（5）使用 cache / interleave / prefetch 这些空间换时间的操作。

为便于更好理解这些操作，通过以下示例进行具体说明。

3.1 生成数据集

这里以手工创建一个非常简单的数据集，该数据包含10个样本，每个样本对应一个标签。

import numpy as np
import tensorflow as tf

data = np.array([0.1, 0.4, 0.6, 0.2, 0.8, 0.8, 0.4, 0.9, 0.3, 0.2])
#其中大于0.5的样本为正样本，即标签记为1，否则为0
label = np.array([0, 0, 1, 0, 1, 1, 0, 1, 0, 0])
#可以通过tf.data.Dataset.from_tensor_slices建立数据集
dataset = tf.data.Dataset.from_tensor_slices((data, label))

for x,y in dataset:
print("[{},{}]".format(x.numpy(),y.numpy()))

import numpy as np

import tensorflow as tf

data = np.array([0.1, 0.4, 0.6, 0.2, 0.8, 0.8, 0.4, 0.9, 0.3, 0.2])

#其中大于0.5的样本为正样本，即标签记为1，否则为0

label = np.array([0, 0, 1, 0, 1, 1, 0, 1, 0, 0])

#可以通过tf.data.Dataset.from_tensor_slices建立数据集

dataset = tf.data.Dataset.from_tensor_slices((data, label))

for x,y in dataset:

print("[{},{}]".format(x.numpy(),y.numpy()))

运行结果：
[0.1,0]
[0.4,0]
[0.6,1]
[0.2,0]
[0.8,1]
[0.8,1]
[0.4,0]
[0.9,1]
[0.3,0]
[0.2,0]

3.2 map

map对dataset中每个元素做出来，这里每个元素为[x,y],函数为one-hot
该函数把标签转换为one-hot编码。
介绍map()这一核心函数。该函数的输入参数map_func应为一个函数，在该函数中实现我们需要的对数据的变换。
具体应用场景如图片加载、数据增强、标签one hot化等。下面以one hot化和添加噪声为例具体说明。
one hot化的函数实现如下

def one_hot(x, y):
    if y == 0:
        return x, np.array([1, 0])
    else:
        return x, np.array([0, 1])

dataset = dataset.map(one_hot)
for x,y in dataset_one_hot:
print("[{},{}]".format(x.numpy(),y.numpy()))

def one_hot(x, y):

if y == 0:

return x, np.array([1, 0])

else:

return x, np.array([0, 1])

dataset = dataset.map(one_hot)

for x,y in dataset_one_hot:

print("[{},{}]".format(x.numpy(),y.numpy()))

运行结果
[0.1,[1 0]]
[0.4,[1 0]]
[0.6,[0 1]]
[0.2,[1 0]]
[0.8,[0 1]]
[0.8,[0 1]]
[0.4,[1 0]]
[0.9,[0 1]]
[0.3,[1 0]]
[0.2,[1 0]]

3.3 shuffle

shuffle()是随机打乱样本次序，参数buffer_size建议设为样本数量，过大会浪费内存空间，过小会导致打乱不充分。

dataset1 = dataset.shuffle(buffer_size=10)
for x,y in dataset1:
print("[{},{}]".format(x.numpy(),y.numpy()))

dataset1 = dataset.shuffle(buffer_size=10)

for x,y in dataset1:

print("[{},{}]".format(x.numpy(),y.numpy()))

运行结果
[0.6,[0 1]]
[0.2,[1 0]]
[0.3,[1 0]]
[0.8,[0 1]]
[0.4,[1 0]]
[0.8,[0 1]]
[0.2,[1 0]]
[0.4,[1 0]]
[0.1,[1 0]]
[0.9,[0 1]]

3.4 repeat

使用repeat方法，repeat的功能就是将整个序列或数据集重复多次，完成整个数据集的一次训练是一个epoch，使用repeat(5)就可以将之变成5个epoch 如果直接调用repeat()的话，生成的序列就会无限重复下去，没有结束，因此也不会抛出tf.errors.OutOfRangeError异常。

dataset_repeat = dataset1.repeat(2)
for x,y in dataset_repeat:
print("[{},{}]".format(x.numpy(),y.numpy()))

dataset_repeat = dataset1.repeat(2)

for x,y in dataset_repeat:

print("[{},{}]".format(x.numpy(),y.numpy()))

运行结果
[0.8,[0 1]]
[0.8,[0 1]]
[0.1,[1 0]]
[0.9,[0 1]]
[0.2,[1 0]]
[0.2,[1 0]]
[0.4,[1 0]]
[0.3,[1 0]]
[0.6,[0 1]]
[0.4,[1 0]]
[0.4,[1 0]]
[0.2,[1 0]]
[0.4,[1 0]]
[0.3,[1 0]]
[0.8,[0 1]]
[0.2,[1 0]]
[0.8,[0 1]]
[0.9,[0 1]]
[0.6,[0 1]]
[0.1,[1 0]]

3.5 batch

batch()是使数据集一次获取多个样本

# batch()是使迭代器一次获取多个样本
dataset_batch =dataset_repeat.batch(batch_size=4)  
for x,y in dataset_batch:
print("[{},{}]".format(x.numpy(),y.numpy()))

# batch()是使迭代器一次获取多个样本

dataset_batch =dataset_repeat.batch(batch_size=4)

for x,y in dataset_batch:

print("[{},{}]".format(x.numpy(),y.numpy()))

运行结果
[[0.4 0.3 0.6 0.2],[[1 0]
[1 0]
[0 1]
[1 0]]]
[[0.4 0.2 0.8 0.9],[[1 0]
[1 0]
[0 1]
[0 1]]]
[[0.8 0.1 0.6 0.9],[[0 1]
[1 0]
[0 1]
[0 1]]]
[[0.2 0.2 0.4 0.1],[[1 0]
[1 0]
[1 0]
[1 0]]]
[[0.3 0.8 0.8 0.4],[[1 0]
[0 1]
[0 1]
[1 0]]]

3.6map

map()函数，该函数的输入参数map_func应为一个函数，在该函数中实现我们需要的对数据的变换。具体应用场景如图片加载、数据增强、标签one hot化等。
对数据进行固定形式上的变化，可将函数直接作为参数输入。但是，包含随机信息的数据变化则需要tf.py_function辅助实现，
如数据增强中数据添加随机噪声、图像的随机翻转都属于包含随机信息。

def add_noise(x, y):
    x += np.random.normal(0.0, 1.0)
    return x, y
dataset_add_noise = dataset.map(lambda x, y: tf.py_function(add_noise, inp=[x, y], Tout=[tf.float64, tf.int32]))
for x,y in dataset_add_noise:
print("[{},{}]".format(x.numpy(),y.numpy()))

def add_noise(x, y):

x += np.random.normal(0.0, 1.0)

return x, y

dataset_add_noise = dataset.map(lambda x, y: tf.py_function(add_noise, inp=[x, y], Tout=[tf.float64, tf.int32]))

for x,y in dataset_add_noise:

print("[{},{}]".format(x.numpy(),y.numpy()))

运行结果
[[1.5635917 1.6635917 2.2635917 2.0635917],[[1 0]
[1 0]
[0 1]
[0 1]]]
[[1.29330552 0.89330552 1.39330552 0.79330552],[[0 1]
[1 0]
[0 1]
[1 0]]]
[[-0.04628853 -0.24628853 -0.04628853 -0.24628853],[[1 0]
[1 0]
[1 0]
[1 0]]]
[[-0.46844772 -0.96844772 -1.16844772 -0.66844772],[[0 1]
[1 0]
[1 0]
[0 1]]]
[[0.40181042 0.20181042 0.80181042 0.90181042],[[1 0]
[1 0]
[0 1]
[0 1]]]
在map()函数中，还有个很重要的参数num_parallel_calls，可以将数据加载与变换过程并行到多个CPU线程上。由于python语言本身的全局解释锁，想要实现真正的并行计算是非常困难的，所以这个参数实际上非常实用，通常的使用情景是网络训练时，GPU做模型运算的同时CPU加载数据。还可以直接设置num_parallel_calls=tf.data.experimental.AUTOTUNE，这样会自动设置为最大的可用线程数，可充分利用机器算资源。

AUTOTUNE = tf.data.experimental.AUTOTUNE
dataset_add_noise1 = dataset.map(lambda x, y: tf.py_function(add_noise, inp=[x, y], Tout=[tf.float64, tf.int32]),num_parallel_calls=AUTOTUNE)

1 2	AUTOTUNE = tf.data.experimental.AUTOTUNE dataset_add_noise1 = dataset.map(lambda x, y: tf.py_function(add_noise, inp=[x, y], Tout=[tf.float64, tf.int32]),num_parallel_calls=AUTOTUNE)

3.7 prefetch

prefetch(buffer_size)创建一个Dataset，从源数据集中预提取元素的，注意：examples.prefetch(2) 将预取2个元素（2个示例），
而examples.batch(20).prefetch(2) 将预取2个元素（2个批次，每个20个示例），buffer_size 表示预取时将缓冲的最大元素数，返回 Dataset。
使用prefetch可有效使用读取数据与模型处理之间松耦合。如下图所示

从上图可以看出，使用prefetch函数之后，读取数据与训练数据就可并发处理了，这就大大提升数据处理效率。

4.tf.data读取输入数据

tf.data的架构如下图所示：

从上图可知，Dataset是一个基类，这个类可实例化成迭代器（Iterator），

4.1 Dataset类

Dataset类下有多个子类，常见的有TextLineDataset、tf.data.FixedLengthRecordDataset、TFRecordDataset等，可用使用这些Dateset的子类获取数据，此外，Dataset还有很多方法，如from_tensor_slices、list_files、map、batch、repeat等等，Dataset的这些方法或子类通常用来读取或处理数据，当使用场景有些不同，tf.data常见的读取数据方式有以下几种：
（1）直接从内存中读取（如NumPy数据），tf.data.Dataset.from_tensor_slices()
（2）使用一个 python 生成器 (generator) 初始化，从生成器中读取数据可以使用
tf.data.Dataset.from_generator（）
（3）从 TFrecords格式文件读取数据，可使用tf.data.TFRecordDataset()
（4）读取文本数据，可使用tf.data.TextLineDataset（）
（5）从二进制文件读取数据，可用tf.data.FixedLengthRecordDataset（）
（6）读取cvs数据，可使用tf.data.experimental.make_csv_dataset（）
（7）从文件集中读取数据，可使用tf.data.Dataset.list_files（）

4.2对象Iterator

Iterator是Dataset中迭代方法的实例化，主要对数据进行访问，包括四种迭代方法，单次、可初始化、可重新初始化、可馈送等，可实现对数据集中元素的快速迭代，供模型训练使用。