Attention_lstm-Code

import tensorflow as tf

def attention_3d_block(inputs,TIME_STEPS,SINGLE_ATTENTION_VECTOR):
    # inputs.shape = (batch_size, time_steps, input_dim)
    # inputs = tf.expand_dims(inputs,1)
    input_dim = int(inputs.shape[2])
    a = tf.keras.layers.Permute((2, 1))(inputs)
    a = tf.keras.layers.Reshape((input_dim, TIME_STEPS))(a)  # this line is not useful. It's just to know which dimension is what.
    a = tf.keras.layers.Dense(TIME_STEPS, activation='softmax')(a)
    if SINGLE_ATTENTION_VECTOR:
        a = tf.keras.layers.Lambda(lambda x: tf.keras.backend.mean(x, axis=1), name='dim_reduction')(a)
        a = tf.keras.layers.RepeatVector(input_dim)(a)
    a_probs = tf.keras.layers.Permute((2, 1), name='attention_vec')(a)
    output_attention_mul = tf.keras.layers.Multiply()([inputs, a_probs])
    return output_attention_mul



def attention_lstm(TIME_STEPS, INPUT_DIM,lstm_units = 32):
    tf.keras.backend.clear_session()  # 清除之前的模型，省得压满内存
    inputs = tf.keras.Input(shape=(TIME_STEPS, INPUT_DIM,))
    x = tf.keras.layers.LSTM(lstm_units, return_sequences=True,dropout=0.5)(inputs)
    x = tf.keras.layers.LSTM(lstm_units, return_sequences=True)(x)
    attention_mul = attention_3d_block(x,TIME_STEPS,1)
    lstm_out = tf.keras.layers.LSTM(lstm_units,recurrent_regularizer=tf.keras.regularizers.l2())(attention_mul)
    attention_mul = tf.keras.layers.Flatten()(lstm_out)
    output = tf.keras.layers.Dense(1)(attention_mul)
    model = tf.keras.Model(inputs=[inputs], outputs=output)
    return model