tf.keras.layers.AdditiveAttentio

加法注意力层, 又名Bahdanau-style attention

tf.keras.layers.AdditiveAttention(
    use_scale=True, **kwargs
)

query' shape: [batch_size, Tq, dim], value's shape: [batch_size, Tv, dim], key's shape: [batch_size, Tv, dim], 计算的步骤如下:

1. 把query和value的shape分别转换成[batch_size, Tq, 1, dim]和[batch_size, 1, Tv, dim]
2. 计算注意力分数[batch_size, Tq, Tv]: scores = tf.reduce_sum(tf.tanh(query + value), axis=-1)
3. 进行softmax: distribution = tf.nn.softmax(scores)
4. 对value加权求和: tf.matmul(distribution, value), 得到shape为batch_size, Tq, dim]的输出

调用参数:

inputs:

• query: [batch_size, Tq, dim]
• value: [batch_size, Tv, dim]
• key: [batch_size, Tv, dim], 如果没有给定, 则默认key=value

mask:

• query_mask: [batch_size, Tq], 如果给定, mask==False的位置输出为0.
• value_mask: [batch_size, Tv], 如果给定, mask==False的位置不会对输出产生贡献.

training: 是否启用dropout

示例:

# Variable-length int sequences.
query_input = tf.keras.Input(shape=(None,), dtype='int32')
value_input = tf.keras.Input(shape=(None,), dtype='int32')

# Embedding lookup.
token_embedding = tf.keras.layers.Embedding(max_tokens, dimension)
# Query embeddings of shape [batch_size, Tq, dimension].
query_embeddings = token_embedding(query_input)
# Value embeddings of shape [batch_size, Tv, dimension].
value_embeddings = token_embedding(value_input)

# CNN layer.
cnn_layer = tf.keras.layers.Conv1D(
    filters=100,
    kernel_size=4,
    # Use 'same' padding so outputs have the same shape as inputs.
    padding='same')
# Query encoding of shape [batch_size, Tq, filters].
query_seq_encoding = cnn_layer(query_embeddings)
# Value encoding of shape [batch_size, Tv, filters].
value_seq_encoding = cnn_layer(value_embeddings)

# Query-value attention of shape [batch_size, Tq, filters].
query_value_attention_seq = tf.keras.layers.AdditiveAttention()(
    [query_seq_encoding, value_seq_encoding])

# Reduce over the sequence axis to produce encodings of shape
# [batch_size, filters].
query_encoding = tf.keras.layers.GlobalAveragePooling1D()(
    query_seq_encoding)
query_value_attention = tf.keras.layers.GlobalAveragePooling1D()(
    query_value_attention_seq)

# Concatenate query and document encodings to produce a DNN input layer.
input_layer = tf.keras.layers.Concatenate()(
    [query_encoding, query_value_attention])

# Add DNN layers, and create Model.
# ...