View source on GitHub
|
Dynamic Embedding is designed for Large-scale Sparse Weights Training.
The file will be introduced as shadow_opsunderdynamic_embedding. It is a submodule of dynamic_embedding`.
In TensorFlow 2.x, tf.function is introduced to speedup the computation.
And also modular programming based on tf.Module
are recommended because of the Pythonic style APIs. But APIs like
embedding_lookup, embedding_lookup_unique, embedding_lookup_sparse, and
safe_embedding_lookup_sparse in dynamic_embedding, are wrappers of
embedding_lookup. And it will create a TrainableWrapper object inside
the function, which doesn't meet the requirements of
tf.function
The shadow_ops submodule is designed to support usage on tf.function
and modular style development, like keras.
class ShadowVariable: ShadowVariable is a eager persistent twin of TrainableWrapper.
embedding_lookup(...): Shadow version of dynamic_embedding.embedding_lookup. It use existed shadow