用于多类文本分类的具有自注意力的 LSTM

Question

I am following the self attention in Keras in the following link How to add attention layer to a Bi-LSTM

I want to apply BI LSTM for multi class text classification with 3 classes.

I try o apply the attention in my code, but I got the error below, how can I solve this problem? can anyone help me please?

Incompatible shapes: [100,3] vs. [64,3]
     [[Node: training_1/Adam/gradients/loss_11/dense_14_loss/mul_grad/BroadcastGradientArgs = BroadcastGradientArgs[T=DT_INT32, _class=["loc:@training_1/Adam/gradients/loss_11/dense_14_loss/mul_grad/Reshape_1"], _device="/job:localhost/replica:0/task:0/device:CPU:0"](training_1/Adam/gradients/loss_11/dense_14_loss/mul_grad/Shape, training_1/Adam/gradients/loss_11/dense_14_loss/mul_grad/Shape_1)]]

class attention(Layer):
    
    def __init__(self, return_sequences=False):
        self.return_sequences = return_sequences
        super(attention,self).__init__()
        
    def build(self, input_shape):
        
        self.W=self.add_weight(name="att_weight", shape=(input_shape[-1],1),
                               initializer="normal")
        self.b=self.add_weight(name="att_bias", shape=(input_shape[1],1),
                               initializer="zeros")
        
        super(attention,self).build(input_shape)
        
    def call(self, x):
        
        e = K.tanh(K.dot(x,self.W)+self.b)
        a = K.softmax(e, axis=1)
        output = x*a
        
        if self.return_sequences:
            return output
        
        return K.sum(output, axis=1)




model = Sequential()
model.add(Embedding(17666, 100, input_length=409))
model.add(Bidirectional(LSTM(32, return_sequences=False)))
model.add(attention(return_sequences=True)) # receive 3D and output 2D
model.add(Dropout(0.3))
model.add(Dense(3, activation='softmax'))


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

from keras.callbacks import EarlyStopping
es = EarlyStopping(monitor='val_loss', mode='min', verbose=1, patience=3)
history777=model.fit(x_train, y_train,
 batch_size=100,
 epochs=30,
 validation_data=(x_val, y_val),
 callbacks=[es])

the model:

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
embedding_14 (Embedding)     (None, 409, 100)          1766600   
_________________________________________________________________
bidirectional_14 (Bidirectio (None, 64)                34048     
_________________________________________________________________
attention_14 (attention)     (None, 64)                128       
_________________________________________________________________
dropout_6 (Dropout)          (None, 64)                0         
_________________________________________________________________
dense_14 (Dense)             (None, 3)                 195       
=================================================================
Total params: 1,800,971
Trainable params: 1,800,971
Non-trainable params: 0
____

Answer 1

pay attention to how you set the return_sequence param in the LSTM and attention layers

your output is 2D so the last return sequence must be set to False while the others must be set to True

Your model must be

model = Sequential()
model.add(Embedding(max_words, emb_dim, input_length=max_len))
model.add(Bidirectional(LSTM(32, return_sequences=True))) # return_sequences=True
model.add(attention(return_sequences=False)) # return_sequences=False
model.add(Dropout(0.3))
model.add(Dense(3, activation='softmax'))

here the full example: https://colab.research.google.com/drive/13l5eAHS5uTUsdqyQNm1Dr4JEXg7Fl2Bo?usp=sharing