Release v0.4.1

Updates:

• Rename arguemt of from_pretrained from model_params to override_params
• Add BertForSentenceEmbedding model for embedding extraction
• Add BertForTokenClassification、AlbertForTokenClassification models for token classify tasks
• Add Conditioanl Random Field based models BertCRFForTokenClassification and AlbertCRFForTokenClassification for token classify tasks

Release v0.4.0

Updates:

• Simplify pretrained models loadding
• Add down-stream models, like BertForQuestionAnswering, BertForSequenceClassification

Release v0.3.1

Updates:

• Convert arguments of Bert.call to List

Release v0.3.0

Changes:

• Refactoring modeling to simplify pretrained weights adapting
• Add new signature to export model in SavedModel format

Release v0.2.6

Updates:

• Pass arguments as tuple in call when subclassing keras.Model, which makes it simplier to export model in saved model format

Release v0.2.5

Updates:

• Convert hidden_states and attention_weights of each encoder layer from List to Tensor
• Update adapters to allow arbitrary model name when loading pretrained checkpoints

Release v0.2.4

Updates:

• Expand arguments of call method instead of as a tuple

Release v0.2.3

Updates:

• Add support for skip loading weights from ckpt
• Add tokenizers

Release v0.2.2

Updates

• Optionally return encoder states and attention weights, both False in default.

Release v0.2.1

Updates:

• Refactored the adapter for more flexible use