如何为 biLSTM 层 Keras 设置自定义初始权重?

How to set custom initial weights to biLSTM layer Keras?

我目前正在构建带有注意力的 BiLSTM,并使用 Antlion 算法优化 BiLSTM 层权重。 Antlion 算法在 MATLAB 代码中,我能够集成 Python 和 MATLAB 以接收优化的权重,如下所示:

#LSTM hidden nodes
hidden_nodes=11

import matlab.engine
eng = matlab.engine.start_matlab()
#call optimised_weights.m 
[forward_kernel, backward_kernel,forward_recurrent, backward_recurrent]=eng.optimised_weights(int(hidden_nodes),nargout=4)
eng.quit()

## convert to nparray 
forward_kernel=np.array(forward_kernel)
backward_kernel=np.array(backward_kernel)
forward_recurrent=np.array(forward_recurrent)
backward_recurrent=np.array(backward_recurrent)

我目前在为 BiLSTM 层设置权重和偏差时遇到问题,如下面的模型(未设置自定义初始权重):

class attention(Layer):
    
    def __init__(self, return_sequences=True,**kwargs):
        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)

    def get_config(self):
        # For serialization with 'custom_objects'
        config = super().get_config()
        config['return_sequences'] = self.return_sequences
        return config

model = Sequential()
model.add(Input(shape=(5,1)))
model.add(Bidirectional(LSTM(hidden_nodes, return_sequences=True)))  
model.add(attention(return_sequences=False)) #this is a custom layer...
model.add(Dense(104, activation="sigmoid"))
model.add(Dropout(0.2))
model.add(Dense(1, activation="sigmoid"))

model.compile(optimizer=tf.keras.optimizers.Adam(epsilon=1e-08,learning_rate=0.01),loss='mse')

es = EarlyStopping(monitor='val_loss', mode='min', verbose=2, patience=50)
mc = ModelCheckpoint('model.h5', monitor='val_loss',
                     mode='min', verbose=2, save_best_only=True)

我试过以下方法:

model.add(Bidirectional(LSTM(hidden_nodes, return_sequences=True,
weights=[forward_kernel,forward_recurrent,np.zeros(20,),backward_kernel,backward_recurrent,np.zeros(20,)]))) 

但是一旦模型被编译,权重和偏差就会改变...即使内核、循环和偏差初始化设置为None...

我已经提到这个 link:https://keras.io/api/layers/initializers/ 但无法将其与我的问题联系起来...

如果你们能提供解决此问题的见解,并且我遗漏了任何基本部分,我将不胜感激。如果需要,我很乐意分享更多详细信息。

谢谢!

使用 tf.constant_initializer 将您的自定义权重提供为 np.array。此外,由于您使用的是 Bidirectional 层,因此您需要明确指定具有自定义权重的后向层。

layer = Bidirectional(
    LSTM(
        hidden_nodes,
        return_sequences=True,
        kernel_initializer=tf.constant_initializer(forward_kernel),
        recurrent_initializer=tf.constant_initializer(forward_recurrent),
    ),
    backward_layer=LSTM(
        hidden_nodes,
        return_sequences=True,
        kernel_initializer=tf.constant_initializer(backward_kernel),
        recurrent_initializer=tf.constant_initializer(backward_recurrent),
        go_backwards=True,
    ),
)

注意权重的预期形状。由于层的输入是 (batch, timesteps, features),您的权重应具有以下形状(考虑 LSTM 单元中的 4 个门):

  • 内核:(features, 4*hidden_nodes)
  • 经常性:(hidden_nodes, 4*hidden_nodes)
  • 偏差:(4*hidden_nodes)