1
import os
2
import time
3
import struct
4
import numpy as np
5
import matplotlib.pyplot as plt
6
import cv2
7

8
def load_mnist(path, kind='train'):
9
    # train: train, test: t10k
10
    labels_path = os.path.join(path,'%s-labels.idx1-ubyte'% kind)
11
    images_path = os.path.join(path,'%s-images.idx3-ubyte'% kind)
12
    with open(labels_path, 'rb') as lbpath:
13
        magic, n = struct.unpack('>II',lbpath.read(8))
14
        labels = np.fromfile(lbpath,dtype=np.uint8)
15
    with open(images_path, 'rb') as imgpath:
16
        magic, num, rows, cols = struct.unpack('>IIII',imgpath.read(16))
17
        images = np.fromfile(imgpath,dtype=np.uint8).reshape(len(labels), 784)
18
    return images, labels
19

20
# 读取MNIST数据
21
train_images, train_labels=load_mnist(".", "train")
22
test_images, test_labels=load_mnist(".", "t10k")
23

24
# 显示MNIST数据
25
fig=plt.figure(figsize=(8,8))
26
fig.subplots_adjust(left=0, right=1, bottom=0, top=1, hspace=0.05, wspace=0.05)
27
for i in range(25):
28
    images = np.reshape(test_images[i], [28,28])
29
    ax = fig.add_subplot(5, 5, i+1, xticks=[], yticks=[])
30
    ax.imshow(images,cmap=plt.cm.binary, interpolation='nearest')
31
    ax.text(0,7,str(test_labels[i]))
32
plt.show()

1
# 数据变形
2
train_labels = train_labels.reshape(len(train_labels), 1)
3
test_labels = test_labels.reshape(len(test_labels), 1)
4

5
# 类型转换
6
train_images = train_images.astype(np.float32)
7
train_labels = train_labels.astype(np.float32)
8
test_images = test_images.astype(np.float32)
9
test_labels = test_labels.astype(np.float32)

1
# 训练
2
knn = cv2.ml.KNearest_create()
3
start = time.time()
4
knn.train(train_images, cv2.ml.ROW_SAMPLE, train_labels)
5
end = time.time()
6
print("训练时长：%d ms" % ((end - start) * 1000))

1
# 推理：使用 K 近邻算法分类
2
K = 11
3
start = time.time()
4
ret, inferences, neighbours, dist = knn.findNearest(test_images, K)
5
end = time.time()
6
print("推理时长：%d ms" % ((end - start) * 1000))

1
# 结果统计
2
cnt = np.zeros(shape=(10,2), dtype="uint32")
3
for idx in range(0, 10000):
4
    if np.abs(inferences[idx][0] - test_labels[idx][0]) < 0.1:
5
        digit = int(test_labels[idx][0])
6
        cnt[digit][0] = cnt[digit][0] + 1
7
        cnt[digit][1] = cnt[digit][1] + 1
8
    else:
9
        digit = int(test_labels[idx][0])
10
        cnt[digit][1] = cnt[digit][1] + 1
11

12
for n in range(0, 10):
13
    print("字符%d的准确率为：%3.2f %%" % (n, 100 * cnt[n][0] / cnt[n][1]))
14

15
total = np.sum(cnt, axis=0)
16
print("整体的准确率为： %3.2f %%" % (100 * total[0] / total[1]))