aistudio地址：
https://aistudio.baidu.com/aistudio/projectdetail/1484526
keras的数字图像识别
一、加载数据
MNIST数据集预加载到Keras库中，包括4个Numpy数组。
然后使用pyplot显示其中一个数组的图片
因为每次都需要重新下载，可以先手动下载到本地，然后加载文件
wget https://storage.googleapis.com/tensorflow/tf-keras-datasets/mnist.npz
from keras.datasets import mnist
import numpy as np
# 使用mnist加载数据
# (train_images, train_labels), (test_images, test_labels) = mnist.load_data()
# 使用本地文件加载数据
train_images = np.load("/home/aistudio/work/mnist/x_train.npy", allow_pickle=True)
train_labels = np.load("/home/aistudio/work/mnist/y_train.npy", allow_pickle=True)
test_images = np.load("/home/aistudio/work/mnist/x_test.npy", allow_pickle=True)
test_labels = np.load("/home/aistudio/work/mnist/y_test.npy", allow_pickle=True)
1.1 查看数据
- 图像是28x28 NumPy数组，像素值介于0到255之间。
- 标签是一个整数数组，范围从0到9.
import matplotlib.pyplot as plt
import matplotlib.image as mpimg
print(train_images.shape)
print(train_labels)
print(test_images.shape)
print(test_labels)
# 25 * 25的grid显示125张图片
plt.figure(figsize=(10,10))
for i in range(25):
plt.subplot(5,5,i+1)
plt.xticks([])
plt.yticks([])
plt.grid(False)
plt.imshow(train_images[i], cmap=plt.cm.binary)
plt.xlabel(train_labels[i])
plt.show()
(60000, 28, 28)
[5 0 4 ... 5 6 8]
(10000, 28, 28)
[7 2 1 ... 4 5 6]
output_3_1.png
二、数据预处理
2.1 图片数据三维转二维
# 三维转二维train_images
train_images_re = train_images.reshape((60000, 28 * 28))
test_images_re = test_images.reshape((10000, 28 * 28))
print(train_images_re.shape)
train_images_re = train_images_re.astype('float32') / 255
test_images_re = test_images_re.astype('float32') / 255
(60000, 784)
2.2 标签分类编码
改成one hot编码。
一个二维数组，数字5转成0. 0. 0. 0. 0. 1. 0. 0. 0. 0.，第五个元素为1.
from keras.utils import to_categorical
train_labels_re = to_categorical(train_labels)
test_labels_re = to_categorical(test_labels)
print('原始: ', train_labels)
print('转化后 - one hot: ')
for i in range(10):
print(train_labels_re[i])
原始: [5 0 4 ... 5 6 8]
转化后 - one hot:
[0. 0. 0. 0. 0. 1. 0. 0. 0. 0.]
[1. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[0. 0. 0. 0. 1. 0. 0. 0. 0. 0.]
[0. 1. 0. 0. 0. 0. 0. 0. 0. 0.]
[0. 0. 0. 0. 0. 0. 0. 0. 0. 1.]
[0. 0. 1. 0. 0. 0. 0. 0. 0. 0.]
[0. 1. 0. 0. 0. 0. 0. 0. 0. 0.]
[0. 0. 0. 1. 0. 0. 0. 0. 0. 0.]
[0. 1. 0. 0. 0. 0. 0. 0. 0. 0.]
[0. 0. 0. 0. 1. 0. 0. 0. 0. 0.]
三、构建网络
3.1添加层
from keras import models
from keras import layers
network = models.Sequential()
# 第一层定义
# 输出，第一维大小：512
# 输入，第一维大小：28 * 28
network.add(layers.Dense(512, activation='relu', input_shape=(28 * 28, )))
# 第二层定义
network.add(layers.Dense(10, activation='softmax'))
3.1 编译
添加损失函数、优化器、监控指标
network.compile(
optimizer='rmsprop',
loss='categorical_crossentropy',
metrics=['accuracy']
)
四、拟合模型
network.fit(
train_images_re,
train_labels_re,
epochs=5,
batch_size=128
)
Epoch 1/5
469/469 [==============================] - 16s 33ms/step - loss: 0.4357 - accuracy: 0.87
Epoch 2/5
469/469 [==============================] - 14s 30ms/step - loss: 0.1135 - accuracy: 0.96
Epoch 3/5
469/469 [==============================] - 15s 31ms/step - loss: 0.0691 - accuracy: 0.97
Epoch 4/5
469/469 [==============================] - 15s 33ms/step - loss: 0.0452 - accuracy: 0.98
Epoch 5/5
469/469 [==============================] - 14s 29ms/step - loss: 0.0352 - accuracy: 0.98
<tensorflow.python.keras.callbacks.History at 0x7f8c27af7190>
五、验证模型
精确度：accuracy
损失度：loss
test_loss, test_acc = network.evaluate(test_images_re, test_labels_re)
print('test_loss', test_loss)
print('test_acc', test_acc)
313/313 [==============================] - 1s 2ms/step - loss: 0.0707 - accuracy: 0.97
test_loss 0.07070968300104141
test_acc 0.9790999889373779
六、预测模型
- 使用predict()方法进行预测，返回样本属于每一个类别的概率
- 使用numpy.argmax()方法找到样本以最大概率所属的类别作为样本的预测标签。
import numpy as np
import matplotlib.pyplot as plt
import matplotlib.image as mpimg
predictions = network.predict(test_images_re)
# 显示预测结果
plt.figure(figsize=(10,10))
for i in range(25):
pre_label = np.argmax(predictions[i])
pre_percent = round(predictions[i][np.argmax(predictions[i])] * 100, 2)
plt.subplot(5,5,i+1)
plt.xticks([])
plt.yticks([])
plt.grid(False)
plt.imshow(test_images[i], cmap=plt.cm.binary)
plt.xlabel(str(pre_percent) + '%: ' + str(pre_label))
plt.show()
output_17_0.png
