- Docs of starspace and starspace_dictionary: fix use of {} in itemize items
- Remove compliance.cpp and compliance.h and move the abort/exit statements directly in the cpp files
- Drop C++11 specification in Makevars
- hereby replacing use of std::random_shuffle with the Fisher-Yates Shuffle Algorithm as indicated at https://gallery.rcpp.org/articles/stl-random-shuffle/ - as data is reshuffled, model training will give different results compared to previous versions due to different randomisations of the training data
- Changes in src/Starspace/src/model.cpp (EmbedModel::train). On Windows, no longer use threads as on CRAN that seems to make the package FAIL.
- Fixed a bug in saving models trained with ngrams > 1. Embeddings of hashed buckets were not saved.
- Default of bucket argument to the textspace C++ function is now changed to 100000 instead of 2000000, impacting all models with ngrams > 1. This was done as the embeddings of the buckets are now saved
- as.matrix.textspace now also allows to get the LHS and RHS embeddings
- Make example conditionally on availability of udpipe
- Changes to vignette: specify maxTrainTime in examples to avoid issues at CRAN caused by randomisation of starting values.
- Changes to src/Makevars
- Changes in src/Starspace/src/model.cpp (EmbedModel::train). On Mac, no longer use threads as on CRAN that seems to make the package FAIL (while on Travis it did not - see issue #10)
- fix for embed_sentencespace & embed_docspace & embed_articlespace, sentences should be separated by "\t", not " \t " in order to avoid having "" in the dictionary
- remove the GNU make as part of the SystemRequirements
- Allow to do transfer learning by passing an embedding matrix and keep on training based on that matrix
- Allow to do semi-supervised learning easily with embed_tagspace
- Attributes attached to a model are now also restored when loading a model with starspace_load_model of type 'ruimtehol'
- Add range.textspace to get the range of embedding similarities
- starspace now also returns the loss before training instead of only starting from epoch 1
- Changes to src/Makevars
- Added -pthread in PKG_CPPFLAGS and removed usage of SHLIB_PTHREAD_FLAGS
- Initial release based on STARSPACE-2017-2.