使用categorical_crossentropy时出错

Question

我正在用tensorflow学习深度学习。我做了一个简单的NLP代码，预测给定句子的下一个单词。

model = tf.keras.Sequential()
model.add(Embedding(num,64,input_length = max_len-1))   # we subtract 1 coz we cropped the laste word from X in out data
model.add(Bidirectional(LSTM(32)))
model.add(Dense(num,activation = 'softmax'))


model.compile(optimizer = 'adam',loss = 'categorical_crossentropy',metrics = ['accuracy'])

history = model.fit(X,Y,epochs = 500)

但是，使用categorical_crossentropy会给出以下错误

ValueError: You are passing a target array of shape (453, 1) while using as loss `categorical_crossentropy`. `categorical_crossentropy` expects targets to be binary matrices (1s and 0s) of shape (samples, classes). If your targets are integer classes, you can convert them to the expected format via:

从keras.utils导入to_categorical

y_binary = to_categorical(y_int)

Alternatively, you can use the loss function `sparse_categorical_crossentropy` instead, which does expect integer targets.

有人能解释一下这是什么意思吗?为什么我不能使用绝对的交叉熵损失函数？非常感谢!任何帮助都将不胜感激！

Answer 1

分类交叉熵用于多类分类问题.当您使用"softmax“作为激活时，将在输出层中为每个类提供一个节点。对于每个示例，对应于示例类的节点应该接近一个，其余节点应该接近零。因此，真正的类标签Y需要是一个单热点编码向量.

假设Y中的类标签是0,1,2，……这样的整数。请试试下面的代码。

from keras.utils import to_categorical

model = tf.keras.Sequential()
model.add(Embedding(num,64,input_length = max_len-1))   # we subtract 1 coz we cropped the laste word from X in out data
model.add(Bidirectional(LSTM(32)))
model.add(Dense(num,activation = 'softmax'))


model.compile(optimizer = 'adam',loss = 'categorical_crossentropy',metrics = ['accuracy'])

Y_one_hot=to_categorical(Y) # convert Y into an one-hot vector
history = model.fit(X,Y_one_hot,epochs = 500)  # use Y_one_hot encoding instead of Y

Answer 2

对于提供的答案(由Roohollah提供)，您必须导入to_categorical，如下所示：

from keras.utils.np_utils import to_categorical