All notable changes to this project will be documented in this file.
- Library structure changed to the tasks-oriented design
- Training with meta text (CTRL).
- Global library structure refactored for further expansion.
- Data format has changed. Now no need for intermediate data arrays preparation. Training is performed on raw text files.
- Now data is processed during the training with multiprocessing workers.
- Embeddings now resized with mean embeddings vector (not the random one).
- Added telegram client for text generator service.
- Added unlikelihood candidates loss.