keras-adaptive-softmax

adaptive-softmax implemented in Keras


Keywords
adaptive, keras, softmax
License
MIT
Install
pip install keras-adaptive-softmax==0.9.0

Documentation

Keras Adaptive Softmax

Version License

[中文|English]

Install

pip install keras-adaptive-softmax

Usage

Generally, AdaptiveEmbedding and AdaptiveSoftmax should be used together. AdaptiveEmbedding provides variable length embeddings, while AdaptiveSoftmax calculates the similarities between the outputs and the generated embeddings.

import keras
from keras_adaptive_softmax import AdaptiveEmbedding, AdaptiveSoftmax

input_layer = keras.layers.Input(shape=(None,))
embed_layer = AdaptiveEmbedding(
    input_dim=30,
    output_dim=32,
    cutoffs=[5, 15, 25],
    div_val=2,
    return_embeddings=True,
    return_projections=True,
    mask_zero=True,
)(input_layer)
dense_layer = keras.layers.Dense(
    units=32,
    activation='tanh',
)(embed_layer[0])
softmax_layer = AdaptiveSoftmax(
    input_dim=32,
    output_dim=30,
    cutoffs=[5, 15, 25],
    div_val=2,
    bind_embeddings=True,
    bind_projections=True,
)([dense_layer] + embed_layer[1:])
model = keras.models.Model(inputs=input_layer, outputs=softmax_layer)
model.compile('adam', 'sparse_categorical_crossentropy')
model.summary()

cutoffs and div_val controls the length of embeddings for each token. Suppose we have 30 distinct tokens, in the above example:

  • The lengths of the embeddings of the first 5 tokens are 32
  • The lengths of the embeddings of the next 10 tokens are 16
  • The lengths of the embeddings of the next 10 tokens are 8
  • The lengths of the embeddings of the last 5 tokens are 4