【510】NLP实战系列(七)—— 进阶版(dropout/stacking/BiLSTM)

参考:Bidirectional 层

  进阶版包含以下技术:

  • Recurrent dropout(循环 dropout), a specific, built-in way to use dropout to fight overfitting in recurrent layers.
    • 使用 dropout 正则化的网络需要更长的时间才能完全收敛,因此网络训练轮次要增加为原来的 2 倍。  
  • Stacking recurrent layers(堆叠循环层), to increase the representational power of the network (at the cost of higher computational loads).
    • 在 keras 中逐个堆叠循环层,所有中间层都应该返回完整的输出序列(一个 3D 张量),而不是只返回最后一个时间步的输出。这个可以指定 return_sequences=True 来实现。
  • Bidirectional recurrent layers(双向循环层), which presents the same information to a recurrent network in different ways, increasing accuracy and mitigating forgetting issues.
    • 需要使用 Bidirectional 层,它的第一个参数是一个循环层实例。Bidirectional 层对这个循环层创建了第二个单独实例,然后使用一个实例按正序处理输入序列,另一个实例按逆序处理输入序列。

1. Bidirectional 层

1.1 语法

keras.layers.Bidirectional(layer, merge_mode='concat', weights=None)

1.2 参数

  • layerRecurrent 实例。
  • merge_mode: 前向和后向 RNN 的输出的结合模式。 为 {'sum', 'mul', 'concat', 'ave', None} 其中之一。 如果是 None,输出不会被结合,而是作为一个列表被返回。

2. 举例

2.1 Recurrent dropout

from keras.layers import LSTM

model = Sequential()
model.add(Embedding(max_features, 32))
model.add(LSTM(32,
               dropout=0.2,
               recurrent_dropout=0.2))
model.add(Dense(1, activation='sigmoid'))

model.compile(optimizer='rmsprop',
              loss='binary_crossentropy',
              metrics=['acc'])
history = model.fit(input_train, y_train,
                    epochs=20,
                    batch_size=128,
                    validation_split=0.2)

  

2.2 Stacking recurrent layers

from keras.layers import LSTM

model = Sequential()
model.add(Embedding(max_features, 32))
model.add(LSTM(32,
               dropout=0.2,
               recurrent_dropout=0.2,
               return_sequences=True))
model.add(LSTM(32,
               dropout=0.2,
               recurrent_dropout=0.2))
model.add(Dense(1, activation='sigmoid'))

model.compile(optimizer='rmsprop',
              loss='binary_crossentropy',
              metrics=['acc'])
history = model.fit(input_train, y_train,
                    epochs=20,
                    batch_size=128,
                    validation_split=0.2)

  

2.3 Bidirectional recurrent layers

from keras.layers import LSTM
from keras.layers import Bidirectional

model = Sequential()
model.add(Embedding(max_features, 32))
model.add(Bidirectional(LSTM(32,
                             dropout=0.2,
                             recurrent_dropout=0.2,
                             return_sequences=True)))
model.add(Bidirectional(LSTM(32,
                             dropout=0.2,
                             recurrent_dropout=0.2)))
model.add(Dense(1, activation='sigmoid'))

model.compile(optimizer='rmsprop',
              loss='binary_crossentropy',
              metrics=['acc'])
history = model.fit(input_train, y_train,
                    epochs=5,
                    batch_size=128,
                    validation_split=0.2)

 

上一篇:a标签什么时候可以设置宽高


下一篇:Fine-Tuning and the Stability of Recurrent Neural Networks