4 网络八股扩展
4.1 自制数据集
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
|
import tensorflow as tf
from PIL import Image
import numpy as np
import os
train_dir = './mnist_image_label/mnist_train_jpg_60000/'
train_txt = './mnist_image_label/mnist_train_jpg_60000.txt'
x_train_savepath = './mnist_image_label/mnist_x_train.npy'
y_train_savepath = './mnist_image_label/mnist_y_train.npy'
test_dir = './mnist_image_label/mnist_test_jpg_10000/'
test_txt = './mnist_image_label/mnist_test_jpg.txt'
x_test_savepath = './mnist_image_label/mnist_x_test.npy'
y_test_savepath = './mnist_image_label/mnist_y_test.npy'
def generateDS(dir, labels_path):
f = open(labels_path, 'r') # 存有 文件名 及其 标签
contents = f.readlines() # 读取所有行
f.close() # 关闭
x, y_ = [], [] # 每张图对应的灰度值数据和标签
for content in contents: # 逐行读出
value = content.split() # 以空格分开
img_path = dir + value[0] # 图片名
img = Image.open(img_path)
img = np.array(img.convert('L')) # 8位灰度图像
img = img/255.
x.append(img) # 灰度值np.array 放入列表
y_.append(value[1]) # 标签 放入列表
print('loading: '+content)
x = np.array(x) # 列表变 np.array
y_ = np.array(y_)
y_ = y_.astype(np.int64) # 标签是64位整型
return x, y_
if os.path.exists(x_train_savepath) and os.path.exists(y_train_savepath) and os.path.exists(x_test_savepath) and os.path.exists(y_test_savepath):
print('---Loading dataset---')
x_train_save = np.load(x_train_savepath) # 读
y_train = np.load(y_train_savepath)
x_test_save = np.load(x_test_savepath)
y_test = np.load(y_test_savepath)
x_train = np.reshape(x_train_save, (len(x_train_save),28,28)) # 变形
x_test = np.reshape(x_test_save, (len(x_test_save),28,28))
else: #不存在，要制作数据集
print('---Generating dataset---')
x_train, y_train = generate(train_dir, train_txt)
x_test, y_test = generate(test_dir, test_txt)
print('---Saving dataset') # 保存为以后使用
x_train_save = np.reshape(x_train, (len(x_train), -1))
x_test_save = np.reshape(x_test, (len(x_test), -1))
np.save(x_train_savepath, x_train_save)
np.save(y_train_savepath, y_train)
np.save(x_test_savepath, x_test_save)
np.save(y_test_savepath, y_test)
|
4.2 数据增强
用于扩展数据集，对图像的增强就是对图像的简单形变，用来应对因拍照角度不同引起的图片变形
1
2
3
4
5
6
7
8
9
|
image_gen_train = tf.keras.preprocessing.image.ImageDataGenerator(
rescale = 所有数据将乘以该数值，
rotation_range = 随机旋转角度数范围，
width_shift_range = 随机宽度偏移量，
height_shift_range = 随机高度偏移量，
horizontal_flip = 是否随机水平翻转，
zoom_range = 随机缩放的范围[1-n, 1+n])
image_gen_train.fit(x_train)
|
其中 x_train 需要是四维，需要变形：x_train = x_train.reshape(x_train.shape[0], 28, 28, 1)，把 (60000, 28, 28) -> (60000, 28, 28, 1)
例如：
1
2
3
4
5
6
7
8
9
|
image_gen_train = ImageDataGenerator(
rescale = 1. / 1., # 若为图像，分母为255时，可归至0~1
rotation_range = 45, # 随机45度旋转
width_shift_range=.15, # 宽度偏移
height_shift_range=.15, # 高度偏移
horizontal_flip=False, # 水平翻转
zoom_range = 0.5 # 将图像随机缩放阈量50%
)
image_gen_train.fit(x_train)
|
模型训练也要改，把 model.fit(x_train, y_train, batch_size=32, ...) 改为 model.fit(image_gen_train.flow(x_train, y_train, batch_size=32), ...)
4.3 断点续训, 存取模型
读取模型：load_weights(文件路径)
1
2
3
4
|
checkpoint_save_path = './checkpoint/mnist.ckpt' # 定义文件路径
if os.path.exists(checkpoint_save_path + '.index'): # 存在索引表说明已经保存过参数了
print('---- Loading the model ----')
model.load_weights(checkpoint_save_path)
|
保存模型：
1
2
3
4
5
6
7
8
9
|
cp_callback = tf.keras.callbacks.ModelCheckpoint(
filepath=checkpoint_save_path, # 保存路径
save_weights_only = True, # 是否只保留 weights
save_best_only = True, # 是否只保留最优结果
# 在训练过程中保存，记录到history中
history = model.fit(x_train, y_train, batch_size=32, epochs=5,
validation_data=(x_test, y_test), validation_freq=1,
callbacks=[cp_callback])
|
4.4 提取可训练参数
返回模型中可训练的参数 model.trainable_variables
设置print输出格式：np.set_printoptions(threshold=超过多少省略显示) np.inf 表示无限大
1
2
3
4
5
6
7
8
9
10
|
print(model.trainable_variables) # 把参数直接打印出来
file = open('./weights.txt', 'w') # 把参数存入文件
for v in model.trainable_variables:
file.write(str(v.name) + '\n')
file.write(str(v.shape) + '\n')
file.write(str(v.numpy()) + '\n')
file.close()
|
4.5 acc/locc 可视化
用于查看训练效果。history中记录了训练集的loss和sparse_categorical_accuracy；测试集上的val_loss 和 val_sparse_categorical_accuracy
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
|
acc = history.history['sparse_categorical_accuracy']
val_acc = history.history['val_sparse_categorical_accuracy']
loss = history.history['loss']
val_loss = history.history['val_loss']
plt.subplot(1,2,1)
plt.plot(acc, label='Training Accuracy')
plt.plot(val_acc, label='Validation Accuracy')
plt.title('Training and Validation Accuracy')
plt.legend()
plt.subplot(1,2,2)
plt.plot(loss, label='Training Loss')
plt.plot(val_loss, label='Validation Loss')
plt.title('Training and Validation Loss')
plt.legend()
plt.show()
|
4.6 调用模型
返回前向传播计算结果：predict(输入特征，batch_size=整数)
三步：复现模型，加载参数，预测
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
|
from PIL import Image
import numpy as np
import tensorflow as tf
model_save_path = './checkpoint/mnist.ckpt'
model = tf.keras.models.Sequential([ # 复现网络
tf.keras.layers.Flatten(),
tf.keras.layers.Dense(128, activation='relu'),
tf.keras.layers.Dense(10, activation='softmax')])
model.load_weights(model_save_path) # 加载网络
preNum = int(input("Input the number of test pictures: ")) # 接收用户输入
for i in range(preNum):
image_path = input("the path of test picture: ") # 输入 1.png
img = Image.open(image_path) # 读取图片
# 预处理输入图片的格式 与 训练数据一致
img = Img.resize((28, 28), Image.ANTIALIAS) # 变为 28x28 与训练图片尺寸相同
img_arr = np.array(img.convert('L')) # 变为灰度图
img_arr = 255 - img_arr # 黑白反转
# 预处理或者变为只有黑白像素的高对比度图片, 2种方法二选一
for i in range(28):
for j in range(28):
if img_arr[i][j] < 200: # 小于200 全变白. 可滤去背景噪声，当阈值选择合适识别效果更好
img_arr[i][j] = 255
else:
img_arr[i][j] = 0
img_arr = img_arr / 255.0 # 归一化
x_predict = img_arr[tf.newaxis, ...] # (28, 28) -> (1, 28, 28)
result = model.predict(x_predict) # 输入网络
pred = tf.argmax(result, axis=1) # 返回最大概率值的索引
print('\n')
tf.print(pred)
|