Keras Self-Attention

[中文|English]

Attention mechanism for processing sequential data that considers the context for each timestamp.

Install

pip install keras-self-attention

Usage

Basic

By default, the attention layer uses additive attention and considers the whole context while calculating the relevance. The following code creates an attention layer that follows the equations in the first section (attention_activation is the activation function of e_{t, t'}):

from tensorflow import keras
from keras_self_attention import SeqSelfAttention


model = keras.models.Sequential()
model.add(keras.layers.Embedding(input_dim=10000,
                                 output_dim=300,
                                 mask_zero=True))
model.add(keras.layers.Bidirectional(keras.layers.LSTM(units=128,
                                                       return_sequences=True)))
model.add(SeqSelfAttention(attention_activation='sigmoid'))
model.add(keras.layers.Dense(units=5))
model.compile(
    optimizer='adam',
    loss='categorical_crossentropy',
    metrics=['categorical_accuracy'],
)
model.summary()

Local Attention

The global context may be too broad for one piece of data. The parameter attention_width controls the width of the local context:

from keras_self_attention import SeqSelfAttention

SeqSelfAttention(
    attention_width=15,
    attention_activation='sigmoid',
    name='Attention',
)

Multiplicative Attention

You can use multiplicative attention by setting attention_type:

from keras_self_attention import SeqSelfAttention

SeqSelfAttention(
    attention_width=15,
    attention_type=SeqSelfAttention.ATTENTION_TYPE_MUL,
    attention_activation=None,
    kernel_regularizer=keras.regularizers.l2(1e-6),
    use_attention_bias=False,
    name='Attention',
)

Regularizer

To use the regularizer, set attention_regularizer_weight to a positive number:

from tensorflow import keras
from keras_self_attention import SeqSelfAttention

inputs = keras.layers.Input(shape=(None,))
embd = keras.layers.Embedding(input_dim=32,
                              output_dim=16,
                              mask_zero=True)(inputs)
lstm = keras.layers.Bidirectional(keras.layers.LSTM(units=16,
                                                    return_sequences=True))(embd)
att = SeqSelfAttention(attention_type=SeqSelfAttention.ATTENTION_TYPE_MUL,
                       kernel_regularizer=keras.regularizers.l2(1e-4),
                       bias_regularizer=keras.regularizers.l1(1e-4),
                       attention_regularizer_weight=1e-4,
                       name='Attention')(lstm)
dense = keras.layers.Dense(units=5, name='Dense')(att)
model = keras.models.Model(inputs=inputs, outputs=[dense])
model.compile(
    optimizer='adam',
    loss={'Dense': 'sparse_categorical_crossentropy'},
    metrics={'Dense': 'categorical_accuracy'},
)
model.summary(line_length=100)

Load the Model

Make sure to add SeqSelfAttention to custom objects:

from tensorflow import keras

keras.models.load_model(model_path, custom_objects=SeqSelfAttention.get_custom_objects())

History Only

Set history_only to True when only historical data could be used:

SeqSelfAttention(
    attention_width=3,
    history_only=True,
    name='Attention',
)

Multi-Head

Please refer to keras-multi-head.