v0.102.1

Release Notes

New Features

Embedding Sharing

Different EmbeddingFeature layers that are initialized with the same embedding_share="xxxx" parameter will now share underlying sparse structures. Based on the model loading order, only the first loaded layer will perform the actual load/save operations on the sparse features, while subsequent layers with the same share identifier will reuse them. If an empty string "" is passed for this parameter, sharing is ignored for the layer.

Elimination Policy

An elimination_logic parameter has been introduced for the AdaGradOptimizer, allowing control over the elimination strategy for sparse features.

Logic Defination

Fixes

SequenceEmbeddingFeature

Resolved issues related to the usage of the cvm plugin within the SequenceEmbeddingFeature.

Full Changelog: v0.102.0...v0.102.1