Tensorflow 笔记 Ⅻ——再梳理猫狗数据集 Transfer Learning 与 TFRecord

参考

前面已经有多篇讲解了 TensorFlow 的高速文件读取管道 TFRecord 与 迁移学习，今天再炒点不同的冷菜，本文将不在过多讲解函数与用法，如下博客可供参考：
•TensorFlow实现kaggle数据集图像分类Ⅰ——tfrecord数据格式掌握
•TensorFlow实现kaggle数据集图像分类Ⅱ——tfrecord数据格式制作，展示，与训练一体综合
•Tensorflow 笔记 Ⅹ——Transfer Learning for Kaggle Dataset

TFRecord

TFRecord格式数据文件应用

TFRecord 是 TensorFlow 中的数据集中存储格式。将数据集整理成 TFRecord 格式后，TensorFlow 就可以高效地读取和处理这些数据集，从而更高效地进行大规模的模型训练。
TFRecord 可以理解为一系列序列化的 tf.train.Example 元素所组成的列表文件，而每一个 tf.train.Example 又由若干个 tf.train.Feature 的字典组成。形式如下：

[
{	# example 1
	'feature_1':tf.train.Feature,
	...
	'feature_k':tf.train.Feature,
},
...
{	# example N
	'feature_1':tf.train.Feature,
	...
	'feature_k':tf.train.Feature,
}
]

TFRecord格式数据文件处理过程

将形式各样的数据集整理为 TFRecord 格式，可以对数据集中的每个元素进行以下步骤：
1）读取该数据元素到内存；
2）将该元素转换为 tf.train.Example 对象（每一个 tf.train.Example 由若干个 tf.train.Feature 的字典组成，因此需要先建立 Feature 的字典）；
3）将该 tf.train.Example 对象序列化为字符串，并通过一个预先定义的 tf.io.TFRecordWriter 写入 TFRecord 文件。读取 TFRecord 数据可按照以下步骤：
1）通过 tf.data.TFRecordDataset 读入原始的 TFRecord 文件（此时文件中的 tf.train.Example 对象尚未被反序列化），获得一个 tf.data.Dataset 数据集对象；
2）通过 Dataset.map 方法，对该数据集对象中的每一个序列化的 tf.train.Example 字符串执行 tf.io.parse_single_example 函数，从而实现反序列化。

代码实现

前情函数

import tensorflow as tf
import numpy as np


tf.__version__

'2.0.0'

map 函数介绍

map(
    map_func, num_parallel_calls=None
)
map函数的目的是对每一个数据进行 map_func 函数变化，例如 map_func 函数的目的是加一操作，那么 map 函数将让所有值都加一
map 函数具有两个输入，一个是 map_func，这是要求输入一个数据处理的函数，另一个是 num_parallel_calls，是并行处理时需求的参数，一般用 None 或者使用 tf.data.experimental.AUTOTUNE 来让计算机自适应并行计算

a = tf.data.Dataset.range(1, 6)
for data in a:
    print(data.numpy(), end=' ')
print('\n*********')

a_map = a.map(map_func=lambda x: x + 1)
for data in a_map:
    print(data.numpy(), end=' ')

1 2 3 4 5 
*********
2 3 4 5 6 

def func(input_data):
    return 2 * input_data

a = tf.data.Dataset.range(1, 6)
for data in a:
    print(data.numpy(), end=' ')
print('\n*********')

a_map = a.map(func)
for data in a_map:
    print(data.numpy(), end=' ')

1 2 3 4 5 
*********
2 4 6 8 10 

tf.data.Dataset.from_tensor_slices

主要针对小批量数据集

X = np.array([520, 502, 521, 555])
Y = np.array([1, 2, 3, 4])

dataset = tf.data.Dataset.from_tensor_slices((X, Y))

for x, y in dataset:
    print(x.numpy(), y.numpy())

del X; del Y

520 1
502 2
521 3
555 4

Dataset 利用 iter() 显式的创建一个 Python 迭代器并使用 next() 方法获取下一个元素

X = tf.constant([520, 502, 521, 555])
Y = tf.constant([1, 2, 3, 4])

dataset = tf.data.Dataset.from_tensor_slices((X, Y))

data_it = iter(dataset)

for i in range(len(X)):
    x, y = data_it.next()
    print(x.numpy(), y.numpy())

del X; del Y

520 1
502 2
521 3
555 4

正式开始

模块导入

import tensorflow as tf
import os
import numpy as np
import matplotlib.pyplot as plt

tf.keras.__version__

'2.2.4-tf'

文件获取函数

获取文件名并对文件名打上相应的标签

def read_image_filenames(data_dir):
    filenames = []
    labels = []
    for file in os.listdir(data_dir):
        filenames.append(os.path.join(data_dir, file))
        name = file.split('.')[0]
        if name == 'cat':
            label = 0
        else:
            label = 1
        labels.append(label)
    
    filenames = tf.constant(filenames)
    labels = tf.constant(labels)
    return filenames, labels

图像解码函数

将图像变换为矩阵，并 resize 与 归一化

def decode_image_and_resize(filename, label):
    image_string = tf.io.read_file(filename)
    image_decoded = tf.image.decode_jpeg(image_string)
    
    image_resized = tf.image.resize(image_decoded, [224, 224]) / 255.0
    
    return image_resized, label

展示读取文件与可视化

train_data_dir = './data/train1/'
filenames, labels = read_image_filenames(train_data_dir)
dataset = tf.data.Dataset.from_tensor_slices((filenames, labels))

demo_dataset = dataset.take(5)

for file, label in demo_dataset:
    print('filename:', file.numpy(), (25-len(file.numpy())) * ' ', 'label:', label.numpy())

del demo_dataset

filename: b'./data/train1/cat.0.jpg'    label: 0
filename: b'./data/train1/cat.1.jpg'    label: 0
filename: b'./data/train1/cat.10.jpg'   label: 0
filename: b'./data/train1/cat.100.jpg'  label: 0
filename: b'./data/train1/cat.1000.jpg'  label: 0

dataset = dataset.map(map_func=decode_image_and_resize,
                      num_parallel_calls=tf.data.experimental.AUTOTUNE)

demo_dataset = dataset.take(5)
fig = plt.gcf()
fig.set_size_inches(20, 4)

i = 0
for file, label in demo_dataset:
    i += 1
    ax = plt.subplot(1, 5, i)
    ax.imshow(file, cmap='binary')
    title = 'label:' + str(label.numpy())
    ax.set_title(title, fontsize=15)
plt.show()

del demo_dataset

函数重构

将前面的操作封装成一个函数，命名为 prepare_dataset

def prepare_dataset(data_dir='data/train1/', buffer_size=10000, batch_size=16):
    filenames, labels = read_image_filenames(data_dir)
    dataset = tf.data.Dataset.from_tensor_slices((filenames, labels))
    dataset = dataset.map(map_func=decode_image_and_resize,
                          num_parallel_calls=tf.data.experimental.AUTOTUNE)
    dataset = dataset.shuffle(buffer_size)
    dataset = dataset.batch(batch_size)
    dataset = dataset.prefetch(tf.data.experimental.AUTOTUNE)
    
    return dataset

dataset_train = prepare_dataset()

demo_dataset = dataset_train.take(1)
fig = plt.gcf()
fig.set_size_inches(20, 20)


for file, label in demo_dataset:
    i = 0
    for i in range(len(label)):
        ax = plt.subplot(4, 4, i+1)
        ax.imshow(file[i], cmap='binary')
        title = 'label:' + str(label.numpy()[i])
        ax.set_title(title, fontsize=20)
        ax.set_xticks([])
        ax.set_yticks([])
        i += 1

plt.show()

建立模型

def vgg16_model(input_shape=(224, 224, 3)):
    vgg16 = tf.keras.applications.vgg16.VGG16(include_top=False,
                                             weights='imagenet', 
                                             input_shape=input_shape)
    for layer in vgg16.layers:
        layer.trainable=False
    
    last = vgg16.output
    
    x = tf.keras.layers.Flatten()(last)
    x = tf.keras.layers.Dense(128, activation='relu')(x)
    x = tf.keras.layers.Dropout(0.3)(x)
    x = tf.keras.layers.Dense(32, activation='relu')(x)
    x = tf.keras.layers.Dropout(0.3)(x)
    x = tf.keras.layers.Dense(2, activation='softmax')(x)
    
    model = tf.keras.models.Model(inputs=vgg16.input, outputs=x)
    model.summary()
    
    return model

摘要输入

通过摘要可以发现，我们只需要训练 300w 参数

model = vgg16_model()

Model: "model"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
input_1 (InputLayer)         [(None, 224, 224, 3)]     0         
_________________________________________________________________
block1_conv1 (Conv2D)        (None, 224, 224, 64)      1792      
_________________________________________________________________
block1_conv2 (Conv2D)        (None, 224, 224, 64)      36928     
_________________________________________________________________
block1_pool (MaxPooling2D)   (None, 112, 112, 64)      0         
_________________________________________________________________
block2_conv1 (Conv2D)        (None, 112, 112, 128)     73856     
_________________________________________________________________
block2_conv2 (Conv2D)        (None, 112, 112, 128)     147584    
_________________________________________________________________
block2_pool (MaxPooling2D)   (None, 56, 56, 128)       0         
_________________________________________________________________
block3_conv1 (Conv2D)        (None, 56, 56, 256)       295168    
_________________________________________________________________
block3_conv2 (Conv2D)        (None, 56, 56, 256)       590080    
_________________________________________________________________
block3_conv3 (Conv2D)        (None, 56, 56, 256)       590080    
_________________________________________________________________
block3_pool (MaxPooling2D)   (None, 28, 28, 256)       0         
_________________________________________________________________
block4_conv1 (Conv2D)        (None, 28, 28, 512)       1180160   
_________________________________________________________________
block4_conv2 (Conv2D)        (None, 28, 28, 512)       2359808   
_________________________________________________________________
block4_conv3 (Conv2D)        (None, 28, 28, 512)       2359808   
_________________________________________________________________
block4_pool (MaxPooling2D)   (None, 14, 14, 512)       0         
_________________________________________________________________
block5_conv1 (Conv2D)        (None, 14, 14, 512)       2359808   
_________________________________________________________________
block5_conv2 (Conv2D)        (None, 14, 14, 512)       2359808   
_________________________________________________________________
block5_conv3 (Conv2D)        (None, 14, 14, 512)       2359808   
_________________________________________________________________
block5_pool (MaxPooling2D)   (None, 7, 7, 512)         0         
_________________________________________________________________
flatten (Flatten)            (None, 25088)             0         
_________________________________________________________________
dense (Dense)                (None, 128)               3211392   
_________________________________________________________________
dropout (Dropout)            (None, 128)               0         
_________________________________________________________________
dense_1 (Dense)              (None, 32)                4128      
_________________________________________________________________
dropout_1 (Dropout)          (None, 32)                0         
_________________________________________________________________
dense_2 (Dense)              (None, 2)                 66        
=================================================================
Total params: 17,930,274
Trainable params: 3,215,586
Non-trainable params: 14,714,688
_________________________________________________________________

开始训练

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

train_epochs = 5

train_history = model.fit(dataset_train, epochs=train_epochs, verbose=1)

Epoch 1/5
625/625 [==============================] - 200s 321ms/step - loss: 0.4188 - accuracy: 0.8132
Epoch 2/5
625/625 [==============================] - 188s 301ms/step - loss: 0.2386 - accuracy: 0.9023
Epoch 3/5
625/625 [==============================] - 183s 293ms/step - loss: 0.1999 - accuracy: 0.9196
Epoch 4/5
625/625 [==============================] - 182s 291ms/step - loss: 0.1626 - accuracy: 0.9371
Epoch 5/5
625/625 [==============================] - 181s 290ms/step - loss: 0.1365 - accuracy: 0.9449

模型存储

需要安装 pyyaml，在终端输入 pip install pyyaml，或在 jupyter 输入 !pip install pyyaml

if not os.path.exists('./yaml_models'):
    os.mkdir('./yaml_models')

yaml_string = model.to_yaml()

with open('./yaml_models/cat_vs_dog.yaml', 'w') as model_file:
    model_file.write(yaml_string)

model.save_weights('./yaml_models/cat_vs_dog.h5')

应用模型与可视化

with open('./yaml_models/cat_vs_dog.yaml') as yamlfile:
    loaded_model_yaml = yamlfile.read()

model = tf.keras.models.model_from_yaml(loaded_model_yaml)
model.load_weights('./yaml_models/cat_vs_dog.h5')

随机选取 num 张进行预测

from tensorflow.keras.preprocessing import image

def read_image_files(test_dir='./data/test1', num=16, image_size=(224, 224)):
    image_name_list = []
    image_list = []
    files = os.listdir(test_dir)
    for i in range(num):
        choose = np.random.randint(len(files))
        image_file = os.path.join(test_dir, files[choose])
        files.remove(files[choose])
        img = image.load_img(image_file, target_size=image_size)
        img_array = image.img_to_array(img)
        image_list.append(img_array)
        image_name_list.append(files[choose])
    test_data = np.array(image_list)
    test_data /= 255.0
    print('INFO:%d images have been loaded!' % num)
    return test_data, image_name_list, image_list, num

import math

def predict_show():
    test_data, image_name_list, image_list, num = read_image_files()
    preds = model.predict(test_data)
    result = np.argmax(preds, 1)

    fig = plt.gcf()
    fig.set_size_inches(20, 20)

    i = 0
    for probs in result:
        ax = plt.subplot(math.ceil(num / 4), 4, i+1)
        ax.imshow(image_list[i].astype('uint8'), cmap='binary')
        if probs == 0:
            label = 'cat→'+ image_name_list[i] + ' probs:' + str(preds[i][0])
        else:
            label = 'dog→' + image_name_list[i] + ' probs:' +str(preds[i][1])
        i += 1
        ax.set_title(label, fontsize=15)
        ax.set_xticks([])
        ax.set_yticks([])

    plt.show()

predict_show()

INFO:16 images have been loaded!

TFRecord 制作

定义文件与标签获取函数

def get_image_filenames(data_dir):
    filenames = []
    labels = []
    for file in os.listdir(data_dir):
        filenames.append(os.path.join(data_dir, file))
        name = file.split('.')[0]
        if name == 'cat':
            label = 0
        else:
            label = 1
        labels.append(label)

    return filenames, labels

定义 TFRecord 写入函数

def write_TFRecord_file(filenames, labels, tfrecord_file):
    with tf.io.TFRecordWriter(tfrecord_file) as writer:
        for filename, label in zip(filenames, labels):
            image = open(filename, 'rb').read()
            feature={
                'image': tf.train.Feature(bytes_list=tf.train.BytesList(value=[image])),
                'label': tf.train.Feature(int64_list=tf.train.Int64List(value=[label]))
            }
            example = tf.train.Example(features=tf.train.Features(feature=feature))
            
            writer.write(example.SerializeToString())

开始写入

tfrecord_file = 'data/train1.tfrecords'

if not os.path.isfile(tfrecord_file):
    filenames, labels = get_image_filenames('data/train1')
    write_TFRecord_file(filenames, labels, tfrecord_file)
    print('INFO: TFRecord file has been wrote to', tfrecord_file, '.')
else:
    print('INFO: TFRecord file already exists.')

INFO: TFRecord file already exists.

TFRecord 解码与可视化

定义描述字典

features_description={
   'label': tf.io.FixedLenFeature([], tf.int64),
   'image' : tf.io.FixedLenFeature([], tf.string),
}

def parse_example(example_string):
    features_dict = tf.io.parse_single_example(example_string, features_description)
    features_dict['image'] = tf.io.decode_jpeg(features_dict['image'])
    features_dict['image'] = tf.image.resize(features_dict['image'], [224, 224]) / 255.0
    
    return features_dict['image'], features_dict['label']

定义解码读取函数

def read_TFRecord_file(tfrecord_file):
    raw_dataset = tf.data.TFRecordDataset(tfrecord_file)
    dataset = raw_dataset.map(parse_example)
    
    return dataset

解码与可视化

buffer_size = 10000

batch_size = 16

dataset_train = read_TFRecord_file(tfrecord_file).shuffle(buffer_size).batch(batch_size)

dataset_train

<BatchDataset shapes: ((None, 224, 224, None), (None,)), types: (tf.float32, tf.int64)>

demo_dataset = dataset_train.take(1)
fig = plt.gcf()
fig.set_size_inches(20, 20)


for file, label in demo_dataset:
    i = 0
    for i in range(len(label)):
        ax = plt.subplot(4, 4, i+1)
        ax.imshow(file[i], cmap='binary')
        title = 'label:' + str(label.numpy()[i])
        ax.set_title(title, fontsize=20)
        ax.set_xticks([])
        ax.set_yticks([])
        i += 1

plt.show()

陷入无尽的循环吧

殊途同归，最终利用 TFRecord 得到最初 dataset，照着前面的葫芦就可以开始训练了，这里省略XX字