第四讲 网络八股拓展--自定义数据集加载

Posted 2021-02-20 wbloger

 1 # -*- coding: utf-8 -*-
 2 
 3 import tensorflow as tf
 4 from PIL import Image
 5 import numpy as np
 6 import os
 7 
 8 rt_path = ‘/home/ubuntu/Documents/中国大学MOOCTF笔记2.1共享给所有学习者/class4/MNIST_FC‘
 9 
10 
11 train_path = os.path.join(rt_path, ‘mnist_image_label/mnist_train_jpg_60000/‘)
12 train_txt = os.path.join(rt_path, ‘mnist_image_label/mnist_train_jpg_60000.txt‘)
13 x_train_savepath = os.path.join(rt_path, ‘mnist_image_label/mnist_x_train.npy‘)
14 y_train_savepath = os.path.join(rt_path, ‘mnist_image_label/mnist_y_train.npy‘)
15 
16 test_path = os.path.join(rt_path, ‘mnist_image_label/mnist_test_jpg_10000/‘)
17 test_txt = os.path.join(rt_path, ‘mnist_image_label/mnist_test_jpg_10000.txt‘)
18 x_test_savepath = os.path.join(rt_path, ‘mnist_image_label/mnist_x_test.npy‘)
19 y_test_savepath = os.path.join(rt_path, ‘mnist_image_label/mnist_y_test.npy‘)
20 
21 
22 def generateds(path, txt):
23     with open(txt, ‘r‘) as f:
24         contents = f.readlines()
25     x, y_ = [], []
26     i = 0
27     j = len(contents)
28     for content in contents:
29         value = content.split() # 以空格分开，图片路径为value[0] , 标签为value[1] , 存入列表
30         img_path = path + value[0] # 拼出图片路径和文件名
31         img = Image.open(img_path) # 读入图片
32         img = np.array(img.convert(‘L‘)) # 图片变为8位宽灰度值的np.array格式
33         img = img / 255. # 数据归一化 （实现预处理）
34         x.append(img)
35         y_.append(value[1]) # 标签贴到列表y_
36         print(‘loading %d of %d:‘%(i, j) + content) # 打印状态提示
37         i += 1
38     
39     x = np.array(x)
40     y_ = np.array(y_)
41     y_ = y_.astype(np.int64)
42     return x, y_  # 返回输入特征x，返回标签y_
43 
44 
45 if os.path.exists(x_train_savepath) and os.path.exists(y_train_savepath) and os.path.exists(
46         x_test_savepath) and os.path.exists(y_test_savepath):
47     print(‘--------------------------------Load Datasets------------------------‘)
48     x_train_save = np.load(x_train_savepath)
49     y_train = np.load(y_train_savepath)
50     x_test_save = np.load(x_test_savepath)
51     y_test = np.load(y_test_savepath)
52     x_train = np.reshape(x_train_save, (len(x_train_save), 28, 28))
53     x_test = np.reshape(x_test_save, (len(x_test_save), 28, 28))
54 else:
55     print(‘-------------------------Generate Datasets-----------------------‘)
56     x_train, y_train = generateds(train_path, train_txt)
57     x_test, y_test = generateds(test_path, test_txt)
58     
59     print(‘--------------------Save Datasets--------------------------‘)
60     x_train_save = np.reshape(x_train, (len(x_train), -1))
61     x_test_save = np.reshape(x_test, (len(x_test), -1))
62     np.save(x_train_savepath, x_train_save)
63     np.save(y_train_savepath, y_train)
64     np.save(x_test_savepath, x_test_save)
65     np.save(y_test_savepath, y_test)
66 
67 
68 model = tf.keras.models.Sequential([
69     tf.keras.layers.Flatten(),
70     tf.keras.layers.Dense(128, activation=‘relu‘),
71     tf.keras.layers.Dense(10, activation=‘softmax‘)
72     ])
73 
74 model.compile(optimizer=‘adam‘,
75               loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=False),
76               metrics=[‘sparse_categorical_accuracy‘])
77 
78 model.fit(x_train, y_train, batch_size=32, epochs=5, validation_data=(x_test, y_test), validation_freq=1)
79 model.summary()
80