v0.9.0

New

• Datasets are now also available on HuggingFace: https://huggingface.co/TopicNet.

Fix

• Datasets downloading via dataset manager (that is, fix datasets site URL address; issue, pr).
• Thetaless regularizer behavior for some "extreme modality cases" (issue, pr, related pr).
• Dataset class internals for newer pandas, also "pandas + numpy" collab in some scores (issue1, issue2, pr).
• Update some demo notebooks (issue1, pr1; issue2, pr2).
• Fix some ARTM-master-related-stuff, freeze "ARTM-compatible" protobuf version for easier installation (related issue, pr1, pr2).
• Freeze all other package versions in order to facilitate the work of argeologists in the future (pr, and minor pr-s: pr1, pr2, pr3, pr4).
• "Lick" Readme: add more links, fix some "formatting irregularities" (issue1, issue2, pr1, pr2).
• Add long project description for proper setup (pr), which is now displayed on TopicNet's PyPI page: https://pypi.org/project/topicnet.

Change

• Dummy model now stores all score values (not only the last one; pr).
• Frozen score now can be added to topic model (this allows loading model after incomplete score saving; pr).

v0.8.0

Fixed

• Is the score out of control or not in controller_cube.py
• Separate/multithread mode for cubes in config_parser.py
• TopicNet is finally Mac-installable using Pip 🎉

Changed

• From now on the library is compatible with Python 3.7 or higher (not Python 3.6)

New

• Ability to define a score controller for any custom score
• Experiment's restore_mode: if something happens during the computation, one can resume the process and proceed from last completed cube
• Dataset's dictionary can be altered using recipes
• Thetaless regularizer is now more user-friendly: only dataset is required as input, not n_dw matrix (the notebook is also updated)
• Score's should_compute: scores may be computed not necessarily on every iteration
• Score's precomputed_data: scores may share some data between each other (eg. one score calculates something and makes the result available for other scores)

v0.7.1

• Reworked top_tokens_viewer and top_documnets_viewer
• Added WNTM recipe
• Added dataset_cooc
• Reworked dataset
• Speed up get_possible_modalities
• Reworked write_vw
• Added new regularizer thetaless and demo of its usage and benefits

Datasets, Recipes, Viewers Update and Rework

Various changes, as can be seen in release commit description

v0.6.1

Fixed

Recipes topicnet.cooking_machine.recipes now included in the assembly on PyPi: setup.py updated, project rebuilt and uploaded.

v0.6.0

New

• Added demo notebooks:
  • one: some comparison of TopicNet with Gensim library
  • two: example of analysis of the 20 Newsgroups dataset, more examples of how one can conduct topic modeling with the help of TopicNet and ARTM
  • three: more about 20 Newsgroups dataset analysis before actual topic modeling

Fixed

• Improved top tokens html display by TopTokensViewer
• Fixed TopicNet installation via pip: now all the necessary packages should be installed automatically.
  So the command pip install topicnet should work just fine... for Linux :)

Future Plans

• Add new regularizers
• Add more abilities to control TopicModel's training process — with model.fit() function
• Make the library installable with pip also for Windows and Mac without BigARTM preinstalled

Functionality update

• Enable end-to-end pip install topicnet (without BigARTM preinstalled) for Linux users.

• Adding recipe import into the library

• Support of custom regularizers

• Experimental feature allowing to change regularization coefficients during model training

• Documentation updated

• Various code improvements

v0.4.1

Recipe demo-notebook and minor changes in cubes added.

v0.4.0

Optimized memory consumption
Recipes and queue in the multithreading part updated.

Performance update

Fixed print lengths