Breaking Changes

  • Changed riseapi mode response format (#1585).
  • Removed support for TensorFlow v1.x: removed all TF-based components, removed TF mentions from documentation, default train class replaced with torch_trainer (#1574).
  • TensorFlow-based models were replaced with the PyTorch-based ones, some models were renamed, various models and components were removed.
    • Replaced Models:
    • Removed Models
      • Classifiers, Doc Retrieval, Go-Bot, Neural Morphological Tagging, NER, ODQA, Ranking, Spelling Correction, Context Question Answering (#1523)
      • ASR, TTS (#1526)
      • ELMO (#1533)
      • Chinese Context Question Answering (#1534)
      • Ranking Models (#1537)
      • Go-Bot (#1544)
      • KBQA, Multitask BERT (#1560)
      • Intent Catcher (#1564)
      • Morpho/Syntax Models (#1573)
    • Removed Components
      • Connectors to: Telegram, Microsoft Bot Framework, Amazon Alexa, Yandex Alice (#1548)
      • Various components (#1563)
  • Removed serialization mechanism

Major Features and Improvements

  • Python 3.8/3.9 support (#1525).
  • Added nested configs overwriting mechanism (#1561).
  • Added case-agnostic distil NER for DREAM (#1570).
  • Added DeepPavlov Topics Classifier model (#1584).
  • Added Russian SuperGLUE models (#1577).
  • Added external metrics support (#1546).
  • Added Jupyter Notebook support to documentation (#1592).

Bug Fixes and Other Changes

  • Requirements updated (#1578).
  • Models deprecation mechanism (#1547).
  • Uploaded DeepPavlov BERT models with MLM & NSP heads parameters (#1502).
  • Fixed en_core_web_sm loading error (#1524)
  • Fixed NER models table view (#1529)
  • Removed special version of Transformers library for certain components (#1532)
  • Fixed tests (#1543)
  • Updated library output during model training (#1572)
  • Fixed ConnectionResetError handling in simple_download (#1586)
  • Minor fixes in KBQA models (#1591)
  • Added iterations count and speed output during training (#1593)
  • Fixed datasets version (#1596)