search

tensorflow / community

Stores documents used by the TensorFlow developer community
Apache License 2.0
1.26k stars 576 forks source link

Added DynamicEmbedding RFC #446

Open divyashreepathihalli opened 1 year ago

divyashreepathihalli commented 1 year ago

Added DynamicEmbedding RFC

Mr-Nineteen commented 1 year ago

Dynamic embedding is a very important feature for us.

When training the sorting model that supports scenarios such as search, recommendation, and advertisement, we encountered the following problems:

  1. For the feature selection of the sorting model in the e-commerce search and promotion scenario, the current industry mostly adopts the idea of large-scale discrete IDs. ID features (product IDs, user IDs, brand IDs, etc.) are large and sparse, and the native TF framework is not applicable.
  2. The TensorFlow variable has a fixed size and cannot dynamically increase the ID without restarting training.

With this feature, the main reasons:

  1. Support dynamic scale-out of dynamic embedding features at the TB level.