在 Keras 中使用 one-hot encoding 创建模型

Create model using one - hot encoding in Keras

我正在研究句子分类问题并尝试使用 Keras 来解决。 词汇表中的唯一单词总数为 36。

在这种情况下,总词汇量是[W1,W2,W3....W36]

所以,如果我有一个句子 [W1 W2 W6 W7 W9],如果我对它进行编码,我会得到一个如下所示的 numpy 数组

[[0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 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 0 0 0 0 0 0 0 0 0 0 0 0]
 [0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 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 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0]
 [0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0]]

形状为 (5,36)

我被困在这里了。我生成的全部是 20000 个具有不同形状的 numpy 数组,即 (N,36) 其中 N 是句子中单词的数量。所以,我有 20,000 个句子用于训练,100 个用于测试,所有句子都标有 (1,36) one-hot encoding

我有 x_train、x_test、y_train 和 y_test

x_test 和 y_test 的维度为 (1,36)

谁能告诉我该怎么做?

我做了一些下面的编码

model = Sequential()
model.add(Dense(512, input_shape=(??????))),
model.add(Activation('relu'))
model.add(Dropout(0.5))
model.add(Dense(num_classes))
model.add(Activation('softmax'))
model.compile(loss='categorical_crossentropy',
          optimizer='adam',
          metrics=['accuracy'])

如有任何帮助,我们将不胜感激。

对@putonspectacles 的更新和回应

非常感谢您花时间和精力进行详细的回复。我尝试对您的代码进行一些小的修改,我认为需要完成这些修改才能使代码正常工作。请在下面找到它

num_classes = 5 
max_words = 20
sentences = ["The cat is in the house","The green boy","computer programs are not alive while the children are"]
labels = np.random.randint(0, num_classes, 3)
y = to_categorical(labels, num_classes=num_classes)
words = set(w for sent in sentences for w in sent.split())
word_map = {w : i+1 for (i, w) in enumerate(words)}
#-Changed the below line the inner for loop sent to sent.split()  
sent_ints = [[word_map[w] for w in sent.split()] for sent in sentences]
vocab_size = len(words)
print(vocab_size)
#-changed the below line - the outer for loop sentences to sent_ints
X = np.array([to_categorical(pad_sequences((sent,), max_words),vocab_size+1)  for sent in sent_ints])
print(X)
print(y)
model = Sequential()
model.add(Dense(512, input_shape=(max_words, vocab_size + 1)))
model.add(LSTM(128))
model.add(Dense(5, activation='softmax'))
model.compile(loss='categorical_crossentropy',
      optimizer='adam',
      metrics=['accuracy'])
model.fit(X,y)

如果没有这些更改,代码将无法工作。当我 运行 上面的代码时,我得到正确的嵌入打印如下

[[[[1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 1.]
[0. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 1. 0. 0.]
[0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 1. 0. 0. 0.]
[0. 0. 0. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 0. 0.]]]


[[[1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[0. 0. 0. 0. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 0.]]]


 [[[1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
 [1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
 [1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
 [1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
 [1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
 [1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
 [1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
 [1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
 [1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
 [1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
 [1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
 [0. 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. 0. 0. 0. 0. 0.]
 [0. 1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
 [0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 1. 0. 0. 0. 0. 0.]
 [0. 0. 0. 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. 0. 0.]
 [0. 0. 0. 0. 0. 0. 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. 0. 0. 0. 0. 0. 0.]]]]



[[0. 0. 0. 0. 1.]
[1. 0. 0. 0. 0.]
[0. 1. 0. 0. 0.]]

但我得到的错误是“检查输入时出错:预期 dense_44_input 有 3 个维度,但得到形状为 (3, 1, 20, 16) 的数组"

当我将输入形状更改为 model.add(密集(512, input_shape=(None,max_words, vocab_size + 1)))

我收到错误“输入 0 与层 lstm_27 不兼容:预期 ndim=3,发现 ndim=4

我正在努力解决这个问题。如果你能给我一个方向,那就太好了。

我接受了这个答案,因为它回答了嵌入单词的objective。再次感谢。

太棒了,你解决了这个问题。你想对一个句子进行分类。我假设你说我想比 bag-of-words 编码做得更好。您想重视顺序。

然后我们将选择一个新模型 -- an RNN (the LSTM version)。该模型有效地总结了每个单词(按顺序)的重要性,因为它构建了最适合任务的句子表示。

但是我们将不得不以不同的方式处理预处理。为了提高效率(以便我们可以批量处理更多句子,而不是一次处理单个句子),我们希望所有句子都具有 相同 数量的单词。所以我们选择 max_words,比如说 20,然后我们填充较短的句子以达到最大单词数,然后我们将超过 20 个单词的句子剪下来。

Keras 将为此提供帮助。我们将用整数对每个单词进行编码。

from keras.preprocessing.sequence import pad_sequences
from keras.utils import to_categorical
from keras.models import Sequential
from keras.layers import Embedding, Dense, LSTM

num_classes = 5 
max_words = 20
sentences = ["The cat is in the house",
                           "The green boy",
            "computer programs are not alive while the children are"]
labels = np.random.randint(0, num_classes, 3)
y = to_categorical(labels, num_classes=num_classes)

words = set(w for sent in sentences for w in sent.split())
word_map = {w : i+1 for (i, w) in enumerate(words)}
sent_ints = [[word_map[w] for w in sent] for sent in sentences]
vocab_size = len(words)

所以 "the green boy" 现在可能是 [1, 3, 5]。 然后我们用

填充和one-hot编码
# pad to max_words length and encode with len(words) + 1  
# + 1 because we'll reserve 0 add the padding sentinel.
X = np.array([to_categorical(pad_sequences((sent,), max_words),  
       vocab_size + 1)  for sent in sent_ints])
print(X.shape) # (3, 20, 16)

现在到模型:我们将添加一个 Dense 层来转换那些热的 单词到密集向量。然后我们使用一个LSTM来转换词向量 in 句子到一个密集的句子向量。最后,我们将使用 softmax 激活在 类.

上生成概率分布
model = Sequential()
model.add(Dense(512, input_shape=(max_words, vocab_size + 1)))
model.add(LSTM(128))
model.add(Dense(5, activation='softmax'))
model.compile(loss='categorical_crossentropy',
          optimizer='adam',
          metrics=['accuracy'])

应该完成了。然后你可以继续训练。

model.fit(X,y)

编辑:

这一行:

# we need to split the sentences in a words write now it reading every
# letter notice the sent.split() in the correct version below.
sent_ints = [[word_map[w] for w in sent] for sent in sentences]

应该是:

sent_ints = [[word_map[w] for w in sent.split()] for sent in sentences]