Bug fixes: - Update version to 2021.2.1 which was out of sync in latest release 206
2021.2.0
Major changes --- - Upgrade spacy to v3.0 - Add native HuggingFace support (191), re-writting `BertClassifier` using transformers - Disables HDBscan from the possible clustering techniques due to a conflict with the new numpy version (197)
Bug fixes --- - Resolves issues 195 and 198 with thew pip reference resolver, introduced in pip>20.3
2021.1.1
Major changes --- - Upgrade spacy to 2.3, transformers to < 2.9 and spacy-tranformers to 0.6 - Transformers functionality (e.g. `SemanticEquivalenceClassifier` and `SemanticEquivalenceClassifier`) are now enabled automatically with extras 'deep-learning'
Bug fixes --- - 173 Fix BertVectorizer for long sequences
2020.11.1
Bug fixes --- - 179 Dataclasses dependency version conflict coming from spacy-transformers - 177 BertClassifier path for scibert when fine tuning not working
2020.11.0
Models ---
CNN/BiLSTM - predict_proba - threshold param default 0.5 - learning_rate decay param default 1 - run training on multiple GPUs if more than one GPUs available - early_stopping param default false - sparse_y param default false - save and load methods
SemanticEquivalenceClassifier - callbacks param with default tensorboard and minibatch history - dropout_rate param default 0.1 - dropout param default true - batch_norm param default true
KerasVectorizer - accepts gensim compatible word vector names and downloads and caches them
Metrics --- - f1 loss and metric
Bug fixes --- - 157 about saving CNN/BiLSTM with attention layer - 80 that fixes indentation and other doc issues - 134 which allows to use tsne as reducer in clustering
2020.9.0
Models --- * Add param l2, dense_size, attention_heads, metrics, callbacks, feature_approach to cnn and bilstm classifiers * Faster predictions in BertClassifier through the use of spacy's pipe * Adds TextClustering