0.2.1

What's Changed

• Fix saving of reduced models: When the models are saved while being reduced, they now properly set the vocab-size in the config, allowing the used to load the model again, no matter if the context manager is still there or not. @helpmefindaname in #4
• If the embeddings are frozen (not trainable), the reduced embeddings will also be frozen @helpmefindaname in #4

Full Changelog: 0.2.0...0.2.1

0.2.0

What's Changed

• Update Optimizer Parameters by @helpmefindaname in #3

Introduces an optional parameter optimizer to reduce_train_vocab. Which can be used to modify the parameter group to exchange changed embeddings to the new ones. An example usage is the following:

  model = ...
  tokenizer = ...
  optimizer = Adam(model.parameters(), lr=...)
  ...

  with reduce_train_vocab(model=model, tokenizer=tokenizer, texts=get_texts_from_dataset(raw_datasets, key="text"), optimizer=optimizer):
     train_with_optimizer(model, tokenizer, optimizer)

  save_model()  # save model at the end to contain the full vocab again.

Full Changelog: 0.1.8...0.2.0

0.1.8

What's Changed

• Lower dependency requirements to transformers 4.1, torch 1.8 and datasets 2.0 as the package was previously too restrictive

Full Changelog: 0.1.7...0.1.8

0.1.7

Full Changelog: 0.1.6...0.1.7

0.1.0

Initial setup, hello world!

So far support for FastTokenizers, BertTokenizer, RobertaTokenizer and XLMRobertaTokenizer is added.