embedding.md

Module: tfra.dynamic_embedding.keras.layers.embedding

View source on GitHub

Dynamic Embedding is designed for Large-scale Sparse Weights Training.

See Sparse Domain Isolation

Classes

class Embedding: A keras style Embedding layer. The Embedding layer acts same like

class FieldWiseEmbedding: An embedding layer, which feature ids are mapped into fields.

class SquashedEmbedding: The SquashedEmbedding layer allow arbirary input shape of feature ids, and get

Functions

reduce_pooling(...): Default combine_fn for Embedding layer. By assuming input