v2.0.3

Enables training dropout for the Bahdanau attention layers.

v2.0.2

Minor fixes including docstring updates and removal of layer name overrides.

v2.0.1

Breaking changes, See pull requests:

• #13
• #14

v1.2.0

• Breaking changes on ServingModel wrapper class, now with dictionary outputs
• new strings module with SubTokenFinderMask class which performs ragged substring searches and masking
• More flexible encoder/decoder network flow

v1.0.0

• Breaking changes for some layers, some layers deprecated (in README)
• Training loop via /bin/main.py is now deprecated and will be removed in future versions
• Included data is removed in favor of the much richer HuggingFace datasets library
• More flexible API for training auto-encoders
• Leverage HuggingFace tokenizers

v0.4.3

• Improved training performance
• Slimmed down tf.saved_model output for inference
• Better documentation.

v0.2.2

Updates include:

• YAML config for training to avoid long CLI inputs
• More flexible method for handling external training data
• convenience CLI function
• PyYAML dependency added
• some improvements to TensorBoard logging for scalar quantities

v0.1.1-beta

Remove the encoding and decoding masking on the Bahdanau attention during the transformer decode pipeline.

v0.1

Initial release of text2vec. This includes tools for creating attention-based and LSTM-based transformer models for turning sentences into vectors which encode contextual meaning.