Giter VIP home page Giter VIP logo

keras-multi-head's Introduction

Keras Multi-Head

Version License

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.

from tensorflow 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.

from tensorflow 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:

from tensorflow 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:

from tensorflow 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 if only one layer is presented as input. The input layers will be considered as query, key and value when a list is given:

from tensorflow import keras
from keras_multi_head import MultiHeadAttention

input_query = keras.layers.Input(
    shape=(2, 3),
    name='Input-Q',
)
input_key = keras.layers.Input(
    shape=(4, 5),
    name='Input-K',
)
input_value = keras.layers.Input(
    shape=(4, 6),
    name='Input-V',
)
att_layer = MultiHeadAttention(
    head_num=3,
    name='Multi-Head',
)([input_query, input_key, input_value])
model = keras.models.Model(inputs=[input_query, input_key, input_value], outputs=att_layer)
model.compile(
    optimizer='adam',
    loss='mse',
    metrics={},
)
model.summary()

Recommend Projects

  • React photo React

    A declarative, efficient, and flexible JavaScript library for building user interfaces.

  • Vue.js photo Vue.js

    ๐Ÿ–– Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.

  • Typescript photo Typescript

    TypeScript is a superset of JavaScript that compiles to clean JavaScript output.

  • TensorFlow photo TensorFlow

    An Open Source Machine Learning Framework for Everyone

  • Django photo Django

    The Web framework for perfectionists with deadlines.

  • D3 photo D3

    Bring data to life with SVG, Canvas and HTML. ๐Ÿ“Š๐Ÿ“ˆ๐ŸŽ‰

Recommend Topics

  • javascript

    JavaScript (JS) is a lightweight interpreted programming language with first-class functions.

  • web

    Some thing interesting about web. New door for the world.

  • server

    A server is a program made to process requests and deliver data to clients.

  • Machine learning

    Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.

  • Game

    Some thing interesting about game, make everyone happy.

Recommend Org

  • Facebook photo Facebook

    We are working to build community through open source technology. NB: members must have two-factor auth.

  • Microsoft photo Microsoft

    Open source projects and samples from Microsoft.

  • Google photo Google

    Google โค๏ธ Open Source for everyone.

  • D3 photo D3

    Data-Driven Documents codes.