Wellcomeml

Release date 01/01/2022

- 379: Fixes a problem in the required dependencies, moving flake8 and black to tests

Release date; 24/11/2021

- 376: Fixes a bug with relative import of EPMC API client

Release date: 05/11/2021

Improvements:
- 328: Expose build_model to CNNClassifier and don't rebuild in fit
- 372: Add more fine grained extras to make vanilla WellcomeML light (109MB)
- 368: Expand LOGGING_LEVEL to TF_CPP_MIN_LOG_LEVEL
- 332: Add `wellcomeml.viz` to vizualize clusters
- 344: Add filter by variable in visualize clusters

Breaking changes:
- 371: Delete `wellcomeml.ml.__init__` so all ml models need to be explicitly imported

Bug fixes:
- 357: Fix sent2vec test error

Release date: 05/08/2021

Bug fixes:

- 346: Fixes problem with tf-idf vectoriser lemmatising twice
- 337: Pin spacy to fix problem with enity-linking test

Release date: 22/07/2021

Improvements:
- 327: Adds verbose and tensorboard_log_path to CNN and SemanticEquivalenceClassifier
- 315: Implements decode to KerasTokenizer and TransformersTokenizer.
- 308: Adds EPMCClient to download data from EPMC
- 283: Disable umap to make imports in wellcomeml faster
- 279: Extend LOGGING_LEVEL env variable to control more libraries logger
- 306: Break down deep-learning extra to tensorflow,torch,spacy for more control over what's installed
- 289: Re-factors frequency vectoriser saving function to make it more efficient

Bug fixes:
- 313: Fix concat feature_approach in CNN
- 297: Fix OOM error in CNN predict
- 292: Fix clustering pipeline when umap used and input length > 4096

Release date: 26/04/2021

Improvements

- 140: Additions to text clustering, exposing each step of the pipeline and adding load/save
- 272: Spacy lemmatiser speed greatly improved
- 255: Improve memory efficiency of TransformersTokenizer
- 245: Voting classifier extras (better input flexibility, parameter for how many models should agree, etc)

- General improvements to continuous testing pipelines

Bug fixes:

- 233: Fix CNN's dataset generator
- 240: Fix multiGPU in SemanticEquivalenceBer
- 237: Fix KerasVectorizer return length