keras-multi-head 0.8.0

A wrapper layer for stacking layers horizontally

A wrapper layer for stacking layers horizontally.

Install

pip install keras-multi-head

Usage

Duplicate Layers

The layer will be duplicated if only a single layer is provided. The layer_num argument controls how many layers will be duplicated eventually.

import keras
from keras_multi_head import MultiHead


model = keras.models.Sequential()
model.add(keras.layers.Embedding(input_dim=100, output_dim=20, name='Embedding'))
model.add(MultiHead(keras.layers.LSTM(units=32), layer_num=5, name='Multi-LSTMs'))
model.add(keras.layers.Flatten(name='Flatten'))
model.add(keras.layers.Dense(units=4, activation='softmax', name='Dense'))
model.build()
model.summary()

Use Multiple-Layers

The first argument could also be a list of layers with different configurations, however, they must have the same output shapes.

import keras
from keras_multi_head import MultiHead


model = keras.models.Sequential()
model.add(keras.layers.Embedding(input_dim=100, output_dim=20, name='Embedding'))
model.add(MultiHead([
    keras.layers.Conv1D(filters=32, kernel_size=3, padding='same'),
    keras.layers.Conv1D(filters=32, kernel_size=5, padding='same'),
    keras.layers.Conv1D(filters=32, kernel_size=7, padding='same'),
], name='Multi-CNNs'))
model.build()
model.summary()

Linear Transformation

The input data will be mapped to different values of the same shape for each layer when hidden_dim is given.

Regularization

The regularization is used when you expect to extract different features from the parallel layers. You can customize the indices of weights in the layers, the intervals represent the parts of the weights and the factor of the regularization.

For example, the bidirectional LSTM layer has 6 weights by default, and the first 3s belong to the forward layer. The 2nd weight (recurrent kernel) in the forward layer controls the computation of gates for recurrent connections. The kernel for computing cell states lays in units x 2 to units x 3 of the recurrent kernel. We can used the regularization for the kernels:

import keras
from keras_multi_head import MultiHead


model = keras.models.Sequential()
model.add(keras.layers.Embedding(input_dim=5, output_dim=3, name='Embed'))
model.add(MultiHead(
    layer=keras.layers.Bidirectional(keras.layers.LSTM(units=16), name='LSTM'),
    layer_num=5,
    reg_index=[1, 4],
    reg_slice=(slice(None, None), slice(32, 48)),
    reg_factor=0.1,
    name='Multi-Head-Attention',
))
model.add(keras.layers.Flatten(name='Flatten'))
model.add(keras.layers.Dense(units=2, activation='softmax', name='Dense'))
model.build()

• reg_index: The indices of layer.get_weights(), a single integer or a list of integers.

• reg_slice: slices or a tuple of slices or a list of the previous choices. If multiple indices are provided in reg_index and reg_slice is not a list, then reg_slice is assumed to be equal for all the indices. The whole array will be used if you leave this argument to None.

• reg_factor: The factor of the regularization, a float or a list of floats.

Multi-Head Attention

A more specific multi-head layer is provided (since the general one is harder to use). The layer uses scaled dot product attention layers as its sub-layers and only head_num is required:

import keras
from keras_multi_head import MultiHeadAttention

input_layer = keras.layers.Input(
    shape=(2, 3),
    name='Input',
)
att_layer = MultiHeadAttention(
    head_num=3,
    name='Multi-Head',
)(input_layer)
model = keras.models.Model(inputs=input_layer, outputs=att_layer)
model.compile(
    optimizer='adam',
    loss='mse',
    metrics={},
)
model.summary()

The shapes of input and output tensors would be the same.