- Pretrained embedding exporter tool to convert embedding .txt or .vec files to tensorflow module in savedmodel format. Also included in Text Cookbook of TF-Hub.
- Minor fixes in Text Classification With Kaggle Notebook. Link here.
Demonstration notebook that uses the embedding exporter tool to classify articles in Bangla using FastText embeddings.
Prototype module that tries to implement ULMFiT technique from Fastai NLP research. It implements a language model module that is trained on Wikitext-103 dataset, then finetunes on IMDB sentiment classification dataset to classify imdb reviews. Still buggy implementation. Details are in the notebooks and scripts included above.
Training loss of language model :
