v0.2.2

Job serialization and clean up.

Changelog

• Minor update providing more flexibility in creating, serializing and saving Jobs.

Full Changelog: v0.2.1...v0.2.2

v0.2.1

What's Changed

• Enable option to train on huggingface datasets. by @harrykeightley in #17

Full Changelog: v0.2.0...v0.2.1

v0.2.0

Major refactor and dependency upgrades.

What's Changed

• Major Refactor by @harrykeightley in #14
  Full Changelog: v0.1.8...v0.2.0

Breaking changes

• elpis.trainer.job.TrainingJob has been removed. This previously contained application logic from elpis_next, forcing you to give a name to your model and dataset, which made no sense to people who might use this as a standalone library.
• Replacing TrainingJob is elpis.models.job.Job, which enables much more flexibility in specifying model, dataset, and training arguments for your job. This Job dataclass is used to store all the training options for the training flow.

Patch transformer versioning errors.

Hotfix for including an incompatible version of transformers with the latest library version which caused errors where the tokenizers had no vocab attribute.

v0.1.7

What's Changed

• Capture tqdm logs by redirecting stderr within context manager by @harrykeightley in #3
• Add Metrics by @harrykeightley in #6
• Build custom Tokenizer and custom Processor flows for wav2vec2 models. by @harrykeightley in #7
• Update poetry version in Github Actions by @harrykeightley in #9
• Dataset handle mismatches by @harrykeightley in #10
• Publish to pip on tagged versions of main. by @harrykeightley in #12

New Contributors

• @harrykeightley made their first contribution in #3

Full Changelog: https://github.com/CoEDL/elpis_lib/commits/v0.1.7