keras 数据集的学习笔记 2/3
上次学习一些常用的数据集，本次将学习数据集的具体应用。
- 各种常用的数据集
- 数据集如何使用
-
定义自己的数据集
深度学习四部曲
- 找数据
- 搭模型
- 做训练
- 搞评测
# 1.找数据
# 考虑到 MNIST的使用的广泛性，我们先从 MNIST数据集入手
# MNIST手写数字识别
%matplotlib inline
from keras.datasets import mnist
import matplotlib.pyplot as plt
(X_train, y_train), (X_test, y_test) = mnist.load_data()
print ('MNIST数据集的形状：', X_train.shape)
print ('训练集的第一个数的向量：', X_train[0])
print ('训练集的第一个数的标签：', y_train[0])
# 可视化数据集
import matplotlib.pyplot as plt
%matplotlib inline
img = X_train[0]
title = y_train[0]
plt.imshow(img)
plt.title(title)
plt.show()
MNIST数据集的形状： (60000, 28, 28)
训练集的第一个数的向量： [[ 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0]
[ 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0]
[ 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0]
[ 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0]
[ 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0]
[ 0 0 0 0 0 0 0 0 0 0 0 0 3 18 18 18 126 136
175 26 166 255 247 127 0 0 0 0]
[ 0 0 0 0 0 0 0 0 30 36 94 154 170 253 253 253 253 253
225 172 253 242 195 64 0 0 0 0]
[ 0 0 0 0 0 0 0 49 238 253 253 253 253 253 253 253 253 251
93 82 82 56 39 0 0 0 0 0]
[ 0 0 0 0 0 0 0 18 219 253 253 253 253 253 198 182 247 241
0 0 0 0 0 0 0 0 0 0]
[ 0 0 0 0 0 0 0 0 80 156 107 253 253 205 11 0 43 154
0 0 0 0 0 0 0 0 0 0]
[ 0 0 0 0 0 0 0 0 0 14 1 154 253 90 0 0 0 0
0 0 0 0 0 0 0 0 0 0]
[ 0 0 0 0 0 0 0 0 0 0 0 139 253 190 2 0 0 0
0 0 0 0 0 0 0 0 0 0]
[ 0 0 0 0 0 0 0 0 0 0 0 11 190 253 70 0 0 0
0 0 0 0 0 0 0 0 0 0]
[ 0 0 0 0 0 0 0 0 0 0 0 0 35 241 225 160 108 1
0 0 0 0 0 0 0 0 0 0]
[ 0 0 0 0 0 0 0 0 0 0 0 0 0 81 240 253 253 119
25 0 0 0 0 0 0 0 0 0]
[ 0 0 0 0 0 0 0 0 0 0 0 0 0 0 45 186 253 253
150 27 0 0 0 0 0 0 0 0]
[ 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 16 93 252
253 187 0 0 0 0 0 0 0 0]
[ 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 249
253 249 64 0 0 0 0 0 0 0]
[ 0 0 0 0 0 0 0 0 0 0 0 0 0 0 46 130 183 253
253 207 2 0 0 0 0 0 0 0]
[ 0 0 0 0 0 0 0 0 0 0 0 0 39 148 229 253 253 253
250 182 0 0 0 0 0 0 0 0]
[ 0 0 0 0 0 0 0 0 0 0 24 114 221 253 253 253 253 201
78 0 0 0 0 0 0 0 0 0]
[ 0 0 0 0 0 0 0 0 23 66 213 253 253 253 253 198 81 2
0 0 0 0 0 0 0 0 0 0]
[ 0 0 0 0 0 0 18 171 219 253 253 253 253 195 80 9 0 0
0 0 0 0 0 0 0 0 0 0]
[ 0 0 0 0 55 172 226 253 253 253 253 244 133 11 0 0 0 0
0 0 0 0 0 0 0 0 0 0]
[ 0 0 0 0 136 253 253 253 212 135 132 16 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0]
[ 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0]
[ 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0]
[ 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0]]
训练集的第一个数的标签： 5
output_2_1.png
这个大家应该很熟悉了吧，上次运行的MNIST数据集, 有60,000个用于训练的28*28的灰度手写数字图片，10,000个测试图片
# 搭模型
# 对于MNIST比较简单的（28*28的灰度图片）数据，我们先做个简单的数据模型吧模型吧
# 2个层的
from keras.models import Sequential
from keras.layers import Dense, Activation
from IPython.display import SVG
from keras.utils.vis_utils import model_to_dot
import pandas as pd
import matplotlib.pyplot as plt
# 定义模型
model = Sequential()
model.add(Dense(32, input_shape=(784,)))
model.add(Activation('relu'))
model.add(Dense(10))
model.add(Activation('softmax'))
# 编译模型
model.compile(optimizer='rmsprop',loss='categorical_crossentropy',metrics=['accuracy'])
# 模型可视化
model.summary()
SVG(model_to_dot(model,show_shapes=True).create(prog='dot', format='svg'))
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
dense_1 (Dense) (None, 32) 25120
_________________________________________________________________
activation_1 (Activation) (None, 32) 0
_________________________________________________________________
dense_2 (Dense) (None, 10) 330
_________________________________________________________________
activation_2 (Activation) (None, 10) 0
=================================================================
Total params: 25,450
Trainable params: 25,450
Non-trainable params: 0
_________________________________________________________________
以上的模型简单也清楚
- 28*28的特征输入 input shape =784
- 全连接层（32）激活函数使用relu
- 全连接层（10）激活函数softmax（用于分类）
看看上图，已经很清楚了。
优化器采用了rmsprop，梯度下降loss采用了categorical_crossentropy，具体一些参数应用可以参考官方的keras的手册。
# 做训练
from keras.utils import np_utils
# 在进行数据处理之前，需要做一下数据的预处理，一般包括数据清洗，归一化等等
X_train = X_train.reshape(X_train.shape[0], -1) / 255. # 归一化
X_test = X_test.reshape(X_test.shape[0], -1) / 255. # 归一化
# 分类标签要做onehot处理
y_train = np_utils.to_categorical(y_train, num_classes=10)
y_test = np_utils.to_categorical(y_test, num_classes=10)
数据预处理是数据进入训练以前必须要做的工作，包括数据清洗，归一化等等
keras内部自带的数据集，像MINIST、CIFAR10等等都是非常标准的数据集，在实际应用中你找不到如此好的数据集，不得不做大量的数据清洗、数据标注等工作，这些都是看起来高大上的深度学习领域中的力气活。
预处理完成后，可以把数据送入编译好的模型中训练。
print('开始训练 ------------')
model.fit(X_train, y_train, epochs=2, batch_size=32)
print('\n测试中 ------------')
loss, accuracy = model.evaluate(X_test, y_test)
print('测试的 loss: ', loss)
print('测试的 accuracy: ', accuracy)
开始训练 ------------
Epoch 1/2
60000/60000 [==============================] - 6s - loss: 0.5595 - acc: 0.8537
Epoch 2/2
60000/60000 [==============================] - 5s - loss: 0.4568 - acc: 0.8777
测试中 ------------
9024/10000 [==========================>...] - ETA: 0s测试的 loss: 0.409912618136
测试的 accuracy: 0.8912
因为minist数据集不大，所以很快就会训练完毕，我们发现测试的losss降到 0.45 正确率到了0.877，貌似还不是很理想。
我们看2个重要的参数 epochs=2, batch_size=32 ，epochs是训练几轮，batch_size是批次的数据数量
我们试图改变一下 把 epochs从2 变成 10看看
print('开始训练 ------------')
model.fit(X_train, y_train, epochs=10, batch_size=32)
print('\n测试中 ------------')
loss, accuracy = model.evaluate(X_test, y_test)
print('测试的 loss: ', loss)
print('测试的 accuracy: ', accuracy)
开始训练 ------------
Epoch 1/10
60000/60000 [==============================] - 5s - loss: 0.4029 - acc: 0.8901
Epoch 2/10
60000/60000 [==============================] - 5s - loss: 0.3708 - acc: 0.8976
Epoch 3/10
60000/60000 [==============================] - 5s - loss: 0.3503 - acc: 0.9018
Epoch 4/10
60000/60000 [==============================] - 5s - loss: 0.3356 - acc: 0.9055
Epoch 5/10
60000/60000 [==============================] - 5s - loss: 0.3246 - acc: 0.9079
Epoch 6/10
60000/60000 [==============================] - 5s - loss: 0.3159 - acc: 0.9095
Epoch 7/10
60000/60000 [==============================] - 5s - loss: 0.3089 - acc: 0.9115
Epoch 8/10
60000/60000 [==============================] - 5s - loss: 0.3028 - acc: 0.9138
Epoch 9/10
60000/60000 [==============================] - 5s - loss: 0.2972 - acc: 0.9148
Epoch 10/10
60000/60000 [==============================] - 5s - loss: 0.2926 - acc: 0.9159
测试中 ------------
9920/10000 [============================>.] - ETA: 0s测试的 loss: 0.283633694214
测试的 accuracy: 0.9171
很棒，我们只是简单的增加了10轮训练，结果loss就降为0.2926，正确率升到 0.9159，不错啊，恭喜一下吧！
那么下一个问题就是，如果我们继续提高 epochs的数值，看看学习效果会不会继续提高呢？
# 为了把学习的效率曲线显示出来，我们这次把loss和accuracy以曲线方式显示，这样很直观
from IPython.display import SVG
from keras.utils.vis_utils import model_to_dot
import pandas as pd
import matplotlib.pyplot as plt
loss = []
acc = []
val_loss = []
val_acc = []
for iteration in range(1, 100):
print('开始训练 ------------')
hist = model.fit(X_train, y_train, epochs=1, batch_size=32, validation_data=(X_test, y_test))
print(hist.history)
loss.append(hist.history['loss'][0])
acc.append(hist.history['acc'][0])
val_loss.append(hist.history['val_loss'][0])
val_acc.append(hist.history['val_acc'][0])
# 图形化整个训练过程
ax = pd.DataFrame(
{
'val_loss': val_loss,
'val_acc': val_acc,
'loss': loss,
'acc': acc,
}
).rolling(5).mean()[5:].plot(title='Training loss', logy=True)
ax.set_xlabel("Epochs")
ax.set_ylabel("Loss&Acc")
开始训练 ------------
Train on 60000 samples, validate on 10000 samples
Epoch 1/1
60000/60000 [==============================] - 7s - loss: 0.2886 - acc: 0.9172 - val_loss: 0.2803 - val_acc: 0.9183
{'val_loss': [0.28033412151634696], 'val_acc': [0.91830000000000001], 'loss': [0.28855849148035051], 'acc': [0.91715000000000002]}
Train on 60000 samples, validate on 10000 samples
Epoch 1/1
60000/60000 [==============================] - 6s - loss: 0.2322 - acc: 0.9344 - val_loss: 0.2335 - val_acc: 0.9340
{'val_loss': [0.23352896861284972], 'val_acc': [0.93400000000000005], 'loss': [0.23224472224811712], 'acc': [0.93438333333333334]}
60000/60000 [==============================] - 6s - loss: 0.2182 - acc: 0.9388 - val_loss: 0.2232 - val_acc: 0.9345
{'val_loss': [0.22320981065034867], 'val_acc': [0.9345], 'loss': [0.21823361753970386], 'acc': [0.9388333333333333]}
开始训练 ------------
Train on 60000 samples, validate on 10000 samples
Epoch 1/1
60000/60000 [==============================] - 6s - loss: 0.2070 - acc: 0.9423 - val_loss: 0.2130 - val_acc: 0.9396
{'val_loss': [0.21298817498832942], 'val_acc': [0.93959999999999999], 'loss': [0.2070259754722317], 'acc': [0.94225000000000003]}
开始训练 ------------
Train on 60000 samples, validate on 10000 samples
Epoch 1/1
60000/60000 [==============================] - 6s - loss: 0.1983 - acc: 0.9444 - val_loss: 0.2076 - val_acc: 0.9414
{'val_loss': [0.20755480074435473], 'val_acc': [0.94140000000000001], 'loss': [0.19831858826875687], 'acc': [0.94443333333333335]}
开始训练 ------------
Train on 60000 samples, validate on 10000 samples
Epoch 1/1
60000/60000 [==============================] - 7s - loss: 0.1970 - acc: 0.9450 - val_loss: 0.2069 - val_acc: 0.9411
{'val_loss': [0.20689115332663061], 'val_acc': [0.94110000000000005], 'loss': [0.1969709436525901], 'acc': [0.94496666666666662]}
开始训练 ------------
Train on 60000 samples, validate on 10000 samples
Epoch 1/1
60000/60000 [==============================] - 6s - loss: 0.1958 - acc: 0.9449 - val_loss: 0.2053 - val_acc: 0.9414
{'val_loss': [0.20527385253384708], 'val_acc': [0.94140000000000001], 'loss': [0.19576028747359911], 'acc': [0.94484999999999997]}
开始训练 ------------
Train on 60000 samples, validate on 10000 samples
Epoch 1/1
60000/60000 [==============================] - 6s - loss: 0.1945 - acc: 0.9459 - val_loss: 0.2029 - val_acc: 0.9424
{'val_loss': [0.20291464587822555], 'val_acc': [0.94240000000000002], 'loss': [0.1944937266851465], 'acc': [0.94589999999999996]}
Epoch 1/1
60000/60000 [==============================] - 6s - loss: 0.1628 - acc: 0.9546 - val_loss: 0.1803 - val_acc: 0.9497
{'val_loss': [0.18027099168971181], 'val_acc': [0.94969999999999999], 'loss': [0.16278639560242494], 'acc': [0.95461666666666667]}
开始训练 ------------
Train on 60000 samples, validate on 10000 samples
Epoch 1/1
60000/60000 [==============================] - 5s - loss: 0.1590 - acc: 0.9553 - val_loss: 0.1798 - val_acc: 0.9498
{'val_loss': [0.17983917704746127], 'val_acc': [0.94979999999999998], 'loss': [0.15900170163338384], 'acc': [0.9553166666666667]}
开始训练 ------------
Train on 60000 samples, validate on 10000 samples
Epoch 1/1
60000/60000 [==============================] - 6s - loss: 0.1589 - acc: 0.9558 - val_loss: 0.1780 - val_acc: 0.9504
{'val_loss': [0.17799708335101605], 'val_acc': [0.95040000000000002], 'loss': [0.15885208675066631], 'acc': [0.95579999999999998]}
开始训练 ------------
Train on 60000 samples, validate on 10000 samples
Epoch 1/1
60000/60000 [==============================] - 6s - loss: 0.1561 - acc: 0.9564 - val_loss: 0.1760 - val_acc: 0.9516
{'val_loss': [0.17602033813893794], 'val_acc': [0.9516], 'loss': [0.15613150059555966], 'acc': [0.95638333333333336]}
开始训练 ------------
Train on 60000 samples, validate on 10000 samples
Epoch 1/1
60000/60000 [==============================] - 6s - loss: 0.1558 - acc: 0.9571 - val_loss: 0.1778 - val_acc: 0.9498
{'val_loss': [0.17784477516561747], 'val_acc': [0.94979999999999998], 'loss': [0.15577643673568964], 'acc': [0.95711666666666662]}
开始训练 ------------
Train on 60000 samples, validate on 10000 samples
Epoch 1/1
60000/60000 [==============================] - 6s - loss: 0.1554 - acc: 0.9564 - val_loss: 0.1758 - val_acc: 0.9514
{'val_loss': [0.17576773921772837], 'val_acc': [0.95140000000000002], 'loss': [0.15542485673725606], 'acc': [0.95640000000000003]}
开始训练 ------------
Train on 60000 samples, validate on 10000 samples
Epoch 1/1
60000/60000 [==============================] - 6s - loss: 0.1523 - acc: 0.9575 - val_loss: 0.1732 - val_acc: 0.9529
{'val_loss': [0.17322715676948428], 'val_acc': [0.95289999999999997], 'loss': [0.15226549789408841], 'acc': [0.95746666666666669]}
开始训练 ------------
Train on 60000 samples, validate on 10000 samples
Epoch 1/1
60000/60000 [==============================] - 6s - loss: 0.1522 - acc: 0.9571 - val_loss: 0.1733 - val_acc: 0.9524
{'val_loss': [0.17329887943789363], 'val_acc': [0.95240000000000002], 'loss': [0.15217291439001759], 'acc': [0.95713333333333328]}
output_12_2.png
# 图形化整个训练过程
ax = pd.DataFrame(
{
'val_loss': val_loss,
'val_acc': val_acc,
'loss': loss,
'acc': acc,
}
).rolling(5).mean()[5:].plot(title='Training loss', logy=True)
ax.set_xlabel("Epochs")
ax.set_ylabel("Loss&Acc")
output_13_1.png
我们发现通过增大训练次数，正确率不断攀升，错误率也是不断下降，只是趋势越来越趋于缓和。
我们再进行100次训练 看看曲线如何？
# 为了把学习的效率曲线显示出来，我们这次把loss和accuracy以曲线方式显示，这样很直观
from IPython.display import SVG
from keras.utils.vis_utils import model_to_dot
import pandas as pd
import matplotlib.pyplot as plt
loss = []
acc = []
val_loss = []
val_acc = []
for iteration in range(1, 100):
print('开始训练 %d： ------------', iteration)
hist = model.fit(X_train, y_train, epochs=1, batch_size=32, validation_data=(X_test, y_test))
#print(hist.history)
loss.append(hist.history['loss'][0])
acc.append(hist.history['acc'][0])
val_loss.append(hist.history['val_loss'][0])
val_acc.append(hist.history['val_acc'][0])
# 图形化整个训练过程
ax = pd.DataFrame(
{
'val_loss': val_loss,
'val_acc': val_acc,
'loss': loss,
'acc': acc,
}
).rolling(5).mean()[5:].plot(title='Training loss', logy=True)
ax.set_xlabel("Epochs")
ax.set_ylabel("Loss&Acc")
开始训练 %d： ------------ 1
Train on 60000 samples, validate on 10000 samples
Epoch 1/1
60000/60000 [==============================] - 7s - loss: 0.1513 - acc: 0.9579 - val_loss: 0.1743 - val_acc: 0.9530
开始训练 %d： ------------ 2
Train on 60000 samples, validate on 10000 samples
Epoch 1/1
60000/60000 [==============================] - 6s - loss: 0.1511 - acc: 0.9576 - val_loss: 0.1741 - val_acc: 0.9527
开始训练 %d： ------------ 3
Train on 60000 samples, validate on 10000 samples
Epoch 1/1
60000/60000 [==============================] - 6s - loss: 0.1505 - acc: 0.9579 - val_loss: 0.1736 - val_acc: 0.9519
开始训练 %d： ------------ 4
Train on 60000 samples, validate on 10000 samples
Epoch 1/1
60000/60000 [==============================] - 6s - loss: 0.1498 - acc: 0.9580 - val_loss: 0.1743 - val_acc: 0.9523
开始训练 %d： ------------ 5
Train on 60000 samples, validate on 10000 samples
Epoch 1/1
60000/60000 [==============================] - 6s - loss: 0.1490 - acc: 0.9587 - val_loss: 0.1716 - val_acc: 0.9529
开始训练 %d： ------------ 6
Train on 60000 samples, validate on 10000 samples
Epoch 1/1
60000/60000 [==============================] - 6s - loss: 0.1486 - acc: 0.9583 - val_loss: 0.1722 - val_acc: 0.9534
开始训练 %d： ------------ 7
Train on 60000 samples, validate on 10000 samples
Epoch 1/1
60000/60000 [==============================] - 6s - loss: 0.1478 - acc: 0.9587 - val_loss: 0.1711 - val_acc: 0.9539
开始训练 %d： ------------ 8
Train on 60000 samples, validate on 10000 samples
Epoch 1/1
60000/60000 [==============================] - 6s - loss: 0.1476 - acc: 0.9590 - val_loss: 0.1713 - val_acc: 0.9533
开始训练 %d： ------------ 9
Train on 60000 samples, validate on 10000 samples
Epoch 1/1
60000/60000 [==============================] - 6s - loss: 0.1472 - acc: 0.9588 - val_loss: 0.1725 - val_acc: 0.9533
开始训练 %d： ------------ 10
Train on 60000 samples, validate on 10000 samples
Epoch 1/1
60000/60000 [==============================] - 6s - loss: 0.1472 - acc: 0.9587 - val_loss: 0.1715 - val_acc: 0.9527
开始训练 %d： ------------ 11
Train on 60000 samples, validate on 10000 samples
Epoch 1/1
60000/60000 [==============================] - 6s - loss: 0.1465 - acc:
Epoch 1/1
60000/60000 [==============================] - 6s - loss: 0.1236 - acc: 0.9664 - val_loss: 0.1611 - val_acc: 0.9606
开始训练 %d： ------------ 98
Train on 60000 samples, validate on 10000 samples
Epoch 1/1
60000/60000 [==============================] - 6s - loss: 0.1233 - acc: 0.9663 - val_loss: 0.1629 - val_acc: 0.9593
开始训练 %d： ------------ 99
Train on 60000 samples, validate on 10000 samples
Epoch 1/1
60000/60000 [==============================] - 6s - loss: 0.1231 - acc: 0.9663 - val_loss: 0.1609 - val_acc: 0.9593
output_15_2.png
我们又做了100次训练，对比上次的100次，发现曲线几乎不再下降了，就是说继续学习下去也不会有太大效果了。深度学习里的术语是：数据不再收敛了。
本次训练的过程：
- 先做2次 loss: 0.4568 - acc: 0.8777 val_loss: 0.409912618136 val_ acc: 0.8912
- 再做10次 loss: 0.2926 - acc: 0.9159 val_loss: 0.283633694214 val acc: 0.9171
- 再做100次 loss: 0.1522 - acc: 0.9571 - val_loss: 0.1733 - val_acc: 0.9524
- 继续做100次 loss: 0.1231 - acc: 0.9663 - val_loss: 0.1609 - val_acc: 0.9593
如果不考虑继续优化模型，我们一般认为，这个时候已经训练完成了，那么我们下一步就是把训练的成果保留住，避免每次使用时都重新训练学习，直接使用训练成果。
from keras.models import Sequential
from keras.layers import Dense, Activation
model.summary()
# 搞评测
# 这种方式也可以用于部署训练好的模型
from keras.models import load_model
# 保存模型和训练的权重
model.save('my_model.h5')
del model # 删除模型
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
dense_1 (Dense) (None, 32) 25120
_________________________________________________________________
activation_1 (Activation) (None, 32) 0
_________________________________________________________________
dense_2 (Dense) (None, 10) 330
_________________________________________________________________
activation_2 (Activation) (None, 10) 0
=================================================================
Total params: 25,450
Trainable params: 25,450
Non-trainable params: 0
_________________________________________________________________
# 加载训练好的模型
model = load_model('my_model.h5')
model.summary()
# 测试一下
print('模型加载后测试: ', model.predict(X_test[0:1]),y_test[0:1])
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
dense_1 (Dense) (None, 32) 25120
_________________________________________________________________
activation_1 (Activation) (None, 32) 0
_________________________________________________________________
dense_2 (Dense) (None, 10) 330
_________________________________________________________________
activation_2 (Activation) (None, 10) 0
=================================================================
Total params: 25,450
Trainable params: 25,450
Non-trainable params: 0
_________________________________________________________________
模型加载后测试: [[ 3.08926704e-11 1.18216120e-17 1.04918041e-09 1.81104348e-03
3.38958772e-09 1.17798572e-06 1.94675519e-20 9.98184144e-01
5.47902061e-07 2.99487897e-06]] [[ 0. 0. 0. 0. 0. 0. 0. 1. 0. 0.]]
以后的代码只加载预训练文件 my_model.h5 就可以了，这个文件包括了模型和训练好的权重。
加载后，就可以使用 model.predict 直接预测结果。我们打印出 X_test[0:1]),y_test[0:1] 看到学习的还不错，都预测对了
