TensorFlow手写数字识别

 

In [1]:

import numpy as np
import matplotlib.pyplot as plt
import tensorflow as tf
from sklearn.preprocessing import StandardScaler
import pandas as pd

 

In [2]:

(x_train_all, y_train_all), (x_test, y_test) = tf.keras.datasets.mnist.load_data()

 

In [3]:

x_valid, x_train = x_train_all[:5000], x_train_all[5000:]
y_valid, y_train = y_train_all[:5000], y_train_all[5000:]

 

In [4]:

scaler = StandardScaler()
x_train.shape

 

Out[4]:

(55000, 28, 28)

 

In [5]:

x_train = x_train.astype(np.float32).reshape(-1, 1)  # 转化成一维的列向量，下面好做标准化（均值和方差好得到）
x_train.shape

 

Out[5]:

(43120000, 1)

 

In [6]:

x_train_scaled = scaler.fit_transform(x_train)
x_train_scaled

 

Out[6]:

array([[-0.4241323],
       [-0.4241323],
       [-0.4241323],
       ...,
       [-0.4241323],
       [-0.4241323],
       [-0.4241323]], dtype=float32)

 

In [7]:

x_train_scaled = x_train_scaled.reshape(-1, 28, 28)   # 做完标准化了，把维度还原回去
display(x_train_scaled.shape)
x_train_scaled

 

 

(55000, 28, 28)

Out[7]:

array([[[-0.4241323, -0.4241323, -0.4241323, ..., -0.4241323,
         -0.4241323, -0.4241323],
        [-0.4241323, -0.4241323, -0.4241323, ..., -0.4241323,
         -0.4241323, -0.4241323],
        [-0.4241323, -0.4241323, -0.4241323, ..., -0.4241323,
         -0.4241323, -0.4241323],
        ...,
        [-0.4241323, -0.4241323, -0.4241323, ..., -0.4241323,
         -0.4241323, -0.4241323],
        [-0.4241323, -0.4241323, -0.4241323, ..., -0.4241323,
         -0.4241323, -0.4241323],
        [-0.4241323, -0.4241323, -0.4241323, ..., -0.4241323,
         -0.4241323, -0.4241323]],
        ...,
        [-0.4241323, -0.4241323, -0.4241323, ..., -0.4241323,
         -0.4241323, -0.4241323],
        [-0.4241323, -0.4241323, -0.4241323, ..., -0.4241323,
         -0.4241323, -0.4241323],
        [-0.4241323, -0.4241323, -0.4241323, ..., -0.4241323,
         -0.4241323, -0.4241323]]], dtype=float32)

 

In [8]:

x_valid_scaled = scaler.transform(x_valid.astype(np.float32).reshape(-1,1)).reshape(-1, 28, 28)
x_test_scaled = scaler.transform(x_test.astype(np.float32).reshape(-1,1)).reshape(-1, 28, 28)

 

In [9]:

model = tf.keras.models.Sequential([
    tf.keras.layers.Flatten(input_shape=[28, 28]),    # 定义输入数据的格式
    tf.keras.layers.Dense(512, activation='relu', kernel_regularizer='l2'),
    # tf.keras.layers.AlphaDropout(0.5),  也可以取消正则化，使用改进版的dropout（均值和方差不变，归一化性质也不变）
    tf.keras.layers.Dense(256, activation='relu', kernel_regularizer='l2'),
    tf.keras.layers.Dense(128, activation='relu', kernel_regularizer='l2'),
    tf.keras.layers.Dense(10, activation='softmax')
])

# 配置模型
model.compile(loss='sparse_categorical_crossentropy', optimizer='sgd', metrics=['accuracy'])

 

In [10]:

history = model.fit(x_train_scaled, y_train, epochs=10, validation_data=(x_valid_scaled, y_valid))

 

 

Epoch 1/10
1719/1719 [==============================] - 6s 3ms/step - loss: 8.6431 - accuracy: 0.9008 - val_loss: 6.0468 - val_accuracy: 0.9378
Epoch 2/10
1719/1719 [==============================] - 5s 3ms/step - loss: 4.4879 - accuracy: 0.9441 - val_loss: 3.2259 - val_accuracy: 0.9522
Epoch 3/10
1719/1719 [==============================] - 5s 3ms/step - loss: 2.4530 - accuracy: 0.9499 - val_loss: 1.8176 - val_accuracy: 0.9588
Epoch 4/10
1719/1719 [==============================] - 5s 3ms/step - loss: 1.4275 - accuracy: 0.9535 - val_loss: 1.1065 - val_accuracy: 0.9584
Epoch 5/10
1719/1719 [==============================] - 5s 3ms/step - loss: 0.9053 - accuracy: 0.9570 - val_loss: 0.7327 - val_accuracy: 0.9618
Epoch 6/10
1719/1719 [==============================] - 5s 3ms/step - loss: 0.6368 - accuracy: 0.9602 - val_loss: 0.5478 - val_accuracy: 0.9636
Epoch 7/10
1719/1719 [==============================] - 5s 3ms/step - loss: 0.4966 - accuracy: 0.9621 - val_loss: 0.4502 - val_accuracy: 0.9652
Epoch 8/10
1719/1719 [==============================] - 5s 3ms/step - loss: 0.4213 - accuracy: 0.9641 - val_loss: 0.3981 - val_accuracy: 0.9632
Epoch 9/10
1719/1719 [==============================] - 5s 3ms/step - loss: 0.3795 - accuracy: 0.9664 - val_loss: 0.3607 - val_accuracy: 0.9688
Epoch 10/10
1719/1719 [==============================] - 5s 3ms/step - loss: 0.3548 - accuracy: 0.9671 - val_loss: 0.3446 - val_accuracy: 0.9684

 

In [12]:

pd.DataFrame(history.history)

	loss	accuracy	val_loss	val_accuracy
0	8.643057	0.900800	6.046839	0.9378
1	4.487876	0.944073	3.225888	0.9522
2	2.453013	0.949891	1.817613	0.9588
3	1.427526	0.953509	1.106460	0.9584
4	0.905319	0.956964	0.732748	0.9618
5	0.636785	0.960236	0.547793	0.9636
6	0.496599	0.962145	0.450182	0.9652
7	0.421305	0.964127	0.398070	0.9632
8	0.379490	0.966382	0.360747	0.9688
9	0.354765	0.967055	0.344589	0.9684

 

In [13]:

pd.DataFrame(history.history).plot(figsize=(8,5))
plt.grid()
plt.gca().set_ylim(0,1)
plt.show()

 

 

 

In [14]:

model.evaluate(x_test_scaled, y_test)

 

 

313/313 [==============================] - 0s 1ms/step - loss: 0.3469 - accuracy: 0.9676

Out[14]:

[0.3468884527683258, 0.9675999879837036]

 

In [15]:

model.summary()

 

 

Model: "sequential"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 flatten (Flatten)           (None, 784)               0                                                               
 dense (Dense)               (None, 512)               401920                                                          
 dense_1 (Dense)             (None, 256)               131328                                                        
 dense_2 (Dense)             (None, 128)               32896                                                           
 dense_3 (Dense)             (None, 10)                1290                                                
=================================================================
Total params: 567434 (2.16 MB)
Trainable params: 567434 (2.16 MB)
Non-trainable params: 0 (0.00 Byte)
_________________________________________________________________