我们可以仅通过整数编码获得有效的结果吗？
Posted
技术标签:
【中文标题】我们可以仅通过整数编码获得有效的结果吗？【英文标题】：Can we achieve effective results with just integer encoding? 【发布时间】：2020-01-11 18:47:11 【问题描述】：我学习深度学习， P>
我使用IMDB数据集。 这是[编码整数]处理？ P>
一些例子显示，你只是在做深度学习，而不是转化为热码一个。 P>
这是足以让一个有效的结果？ P>
如果是那样 P>
什么是热编码之一的优势？ P>
这是我的代码 P>
import numpy
from keras.datasets import imdb
from keras.models import Sequential
from keras.layers import Dense
from keras.layers import Flatten
from keras.layers.convolutional import Conv1D
from keras.layers.convolutional import MaxPooling1D
from keras.layers.embeddings import Embedding
from keras.preprocessing import sequence
# load the dataset but only keep the top n words, zero the rest
top_words = 5000
max_words = 500
X_train = train_result
y_train = train_label
X_test = test_result
y_test = test_label# pad dataset to a maximum review length in words
X_train = sequence.pad_sequences(X_train, maxlen=max_words)
X_test = sequence.pad_sequences(X_test, maxlen=max_words)
print(X_train[:1])
# create the model
model = Sequential()
model.add(Embedding(top_words, 32, input_length=max_words))
model.add(Conv1D(filters=32, kernel_size=3, padding='same', activation='relu'))
model.add(MaxPooling1D(pool_size=2))
model.add(Flatten())
model.add(Dense(250, activation='relu'))
model.add(Dense(1, activation='sigmoid'))
model.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])
model.summary()
# Fit the model
hist = model.fit(X_train, y_train, validation_data=(X_test, y_test), epochs=20, batch_size=128, verbose=1)
# Final evaluation of the model
scores = model.evaluate(X_test, y_test, verbose=0)
print("Accuracy: %.2f%%" % (scores[1]*100))
在X_train [1]。 P>
[[ 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 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 3 284 2452 756 1 3075 194
54 3717 10 757 169 2216 5 1 1906 843 54 52 2732 3403
5 1819 3 34 4 54 1819 5 2532 42 668 23 54 709
52 7 9 2 80 172 3258 265 33 1 1467 4 683 4
11 21 988 1 3 110 631 2 4 321 3 3040 294 284
478 33 1 33 54 4349 33 54 213 2 86 54 516 420
754 1 84 2 8 526 473 63 20 184 20 184 20 184
1138 52 3 23 1 1468 101 3 1850 4 61 6 777 20
237 185 52 3846 5 54 149 7 34 4 1 18 54 4802
929 2 5 98 8 13 17 9 1 993 117 101 3 165
41 653 781 3 286 923 2882 7 210 3 181 5 1743 3
120 814 1630 1517 3 2317 4606 4425 9 43 686 5 744 1018
910 223 136 3782 1585 775 1391 3041 155 3 292 4 2975 2
136 135 120 864 24 869 3655 245 421 1 1803 10 1 120
2 1 261 78 1671 19 43 1288 16 1 1036 5 380 1
1744 121 10 1 84 252 55 51 670 2 24 200 51 1709
1 1256 1469 2 1 217 5 2453 423 79 929 36 9 3
1106 4 2754 4526 14 29 24 2393 74 34 4049 17 42 72
9 365 1 69 41 1804 572 41 559 76 92 2 153 112
11 15 835 1423 136 1 59 15 67 1 1320 5 441 2
733 17 1 688 890 5 26 421 55 23 208 2 31 2070
23 1 2998 136 6 413 44 33 40 7 119 9 668 4
22 3213 40 7 119 151 359 5 25 185]]
，这是输出， P>
Epoch 1/20
10103/10103 [==============================] - 5s 523us/step - loss: 0.5812 - acc: 0.6589 - val_loss: 0.1229 - val_acc: 0.9551
Epoch 2/20
10103/10103 [==============================] - 5s 478us/step - loss: 0.1299 - acc: 0.9485 - val_loss: 0.0693 - val_acc: 0.9663
Epoch 3/20
10103/10103 [==============================] - 5s 488us/step - loss: 0.0544 - acc: 0.9824 - val_loss: 0.0589 - val_acc: 0.9775
Epoch 4/20
10103/10103 [==============================] - 5s 488us/step - loss: 0.0258 - acc: 0.9923 - val_loss: 0.0371 - val_acc: 0.9850
Epoch 5/20
10103/10103 [==============================] - 5s 483us/step - loss: 0.0120 - acc: 0.9976 - val_loss: 0.0528 - val_acc: 0.9813
Epoch 6/20
10103/10103 [==============================] - 5s 483us/step - loss: 0.0058 - acc: 0.9991 - val_loss: 0.0464 - val_acc: 0.9850
Epoch 7/20
10103/10103 [==============================] - 5s 482us/step - loss: 0.0032 - acc: 0.9994 - val_loss: 0.0707 - val_acc: 0.9738
Epoch 8/20
10103/10103 [==============================] - 5s 485us/step - loss: 0.0022 - acc: 0.9997 - val_loss: 0.0471 - val_acc: 0.9925
Epoch 9/20
10103/10103 [==============================] - 5s 482us/step - loss: 0.0011 - acc: 0.9998 - val_loss: 0.0698 - val_acc: 0.9775
Epoch 10/20
10103/10103 [==============================] - 5s 481us/step - loss: 6.8280e-04 - acc: 1.0000 - val_loss: 0.0728 - val_acc: 0.9775
Epoch 11/20
10103/10103 [==============================] - 5s 483us/step - loss: 4.8174e-04 - acc: 1.0000 - val_loss: 0.0873 - val_acc: 0.9738
Epoch 12/20
10103/10103 [==============================] - 5s 477us/step - loss: 3.4037e-04 - acc: 1.0000 - val_loss: 0.0674 - val_acc: 0.9813
Epoch 13/20
10103/10103 [==============================] - 5s 478us/step - loss: 2.6164e-04 - acc: 1.0000 - val_loss: 0.0847 - val_acc: 0.9775
Epoch 14/20
10103/10103 [==============================] - 5s 478us/step - loss: 2.0453e-04 - acc: 1.0000 - val_loss: 0.0812 - val_acc: 0.9775
Epoch 15/20
10103/10103 [==============================] - 5s 473us/step - loss: 1.6034e-04 - acc: 1.0000 - val_loss: 0.0831 - val_acc: 0.9775
Epoch 16/20
10103/10103 [==============================] - 5s 469us/step - loss: 1.3443e-04 - acc: 1.0000 - val_loss: 0.0874 - val_acc: 0.9775
Epoch 17/20
10103/10103 [==============================] - 5s 467us/step - loss: 1.1035e-04 - acc: 1.0000 - val_loss: 0.0891 - val_acc: 0.9775
Epoch 18/20
10103/10103 [==============================] - 5s 471us/step - loss: 9.3257e-05 - acc: 1.0000 - val_loss: 0.0956 - val_acc: 0.9775
Epoch 19/20
10103/10103 [==============================] - 5s 465us/step - loss: 7.9740e-05 - acc: 1.0000 - val_loss: 0.0965 - val_acc: 0.9775
Epoch 20/20
10103/10103 [==============================] - 5s 467us/step - loss: 6.7700e-05 - acc: 1.0000 - val_loss: 0.0919 - val_acc: 0.9775
Accuracy: 97.75%
【问题讨论】：
【参考方案1】：整数编码意味着标签中存在一些序数关系，因此在构建分类模型时需要一次性嵌入。本质上，one-hot embedding 是将离散数据映射到欧几里得空间。
例如，这里的数据集包括 3 个类别：苹果、橙子、香蕉。如果你使用整数编码：apple => 0, orange => 1,banana => 2，你永远不能说'orange'大于或大于'apple'。
在您的情况下，IMDB 评论数据集是一个二元分类数据集，有 2 种标签：负和正。您可以将它们视为连续特征：如果预测值更接近 1，则评论率更高，反之亦然。
https://www.quora.com/What-are-good-ways-to-handle-discrete-and-continuous-inputs-together
Why does one hot encoding improve machine learning performance?
【讨论】：
以上是关于我们可以仅通过整数编码获得有效的结果吗？的主要内容，如果未能解决你的问题，请参考以下文章
我们可以通过 react native 获取当前的壁纸图像吗？
SQL Server 数据库可以通过英国邮政编码驱动 Google 地图吗？