Skip to content
413 lines (329 sloc) 20.1 KB
0.12.5, 2016
* build_vocab takes progress_per parameter for smaller output (@zer0n, #624)
* Control whether to use lowercase for computing word2vec accuracy. (@alantian, #607)
* Easy import of GloVe vectors using Gensim (Manas Ranjan Kar, #625)
- Allow easy port of GloVe vectors into Gensim
- Standalone script with command line arguments, compatible with Python>=2.6
- Usage: python -m gensim.scripts.glove2word2vec -i glove_vectors.txt -o output_word2vec_compatible.txt
* Add `similar_by_word()` and `similar_by_vector()` to word2vec (@isohyt, #381)
0.12.4, 29/01/2016
* Better internal handling of job batching in word2vec (#535)
- up to 300% speed up when training on very short documents (~tweets)
* Word2vec CLI in line with original word2vec.c (Andrey Kutuzov, #538)
- Same default values. See diff
- Standalone script with command line arguments matching those of original C tool.
- Usage: python -m gensim.scripts.word2vec_standalone -train data.txt -output trained_vec.txt -size 200 -window 2 -sample 1e-4
* Improved load_word2vec_format() performance (@svenkreiss, #555)
- Remove `init_sims()` call for performance improvements when normalized vectors are not needed.
- Remove `norm_only` parameter (API change). Call `init_sims(replace=True)` after the `load_word2vec_format()` call for the old `norm_only=True` behavior.
* Word2vec allows non-strict unicode error handling (ignore or replace) (Gordon Mohr, #466)
* Doc2Vec `model.docvecs[key]` now raises KeyError for unknown keys (Gordon Mohr, #520)
* Fix `DocvecsArray.index_to_doctag` so `most_similar()` returns string doctags (Gordon Mohr, #560)
* On-demand loading of the `pattern` library in utils.lemmatize (Jan Zikes, #461)
- `utils.HAS_PATTERN` flag moved to `utils.has_pattern()`
* Threadsafe Word2Vec/Doc2Vec finish-check to avoid hang/unending Word2Vec/Doc2Vec training (Gordon Mohr, #571)
* Tuned `TestWord2VecModel.test_cbow_hs()` against random failures (Gordon Mohr, #531)
* Prevent ZeroDivisionError when `default_timer()` indicate no elapsed time (Gordon Mohr, #518)
* Forwards compatibility for NumPy > 1.10 (Matti Lyra, #494, #513)
- LdaModel and LdaMulticore produce a large number of DeprecationWarnings from
.inference() because the term ids in each chunk returned from utils.grouper
are floats. This behaviour has been changed so that the term IDs are now ints.
- utils.grouper returns a python list instead of a numpy array in .update() when
LdaModel is called in non distributed mode
- in distributed mode .update() will still call utils.grouper with as_numpy=True
to save memory
- LdaModel.update and LdaMulticore.update have a new keyword parameter
chunks_as_numpy=True/False (defaults to False) that allows controlling
this behaviour
0.12.3, 05/11/2015
* Make show_topics return value consistent across models (Christopher Corley, #448)
- All models with the `show_topics` method should return a list of
`(topic_number, topic)` tuples, where `topic` is a list of
`(word, probability)` tuples.
- This is a breaking change that affects users of the `LsiModel`, `LdaModel`,
and `LdaMulticore` that may be reliant on the old tuple layout of
`(probability, word)`.
* Mixed integer & string document-tags (keys to doc-vectors) will work (Gordon Mohr, #491)
- DocvecsArray's `index2doctag` list is renamed/reinterpreted as `offset2doctag`
- `offset2doctag` entries map to `doctag_syn0` indexes *after* last plain-int doctag (if any)
- (If using only string doctags, `offset2doctag` may be interpreted same as `index2doctag`.)
* New Tutorials on Dynamic Topic Modelling and Classification via Word2Vec (@arttii #471, @mataddy #500)
* Auto-learning for the eta parameter on the LdaModel (Christopher Corley, #479)
* Python 3.5 support
* Speed improvements to keyword and summarisation methods (@erbas #441)
* OSX wheels (#504)
* Win build (#492)
0.12.2, 19/09/2015
* tutorial on text summarization (Ólavur Mortensen, #436)
* more flexible vocabulary construction in word2vec & doc2vec (Philipp Dowling, #434)
* added support for sliced TransformedCorpus objects, so that after applying (for instance) TfidfModel the returned corpus remains randomly indexable. (Matti Lyra, #425)
* changed the so that a custom `ignore` list can be passed in (Matti Lyra, #331)
* added support for NumPy style fancy indexing to corpus objects (Matti Lyra, #414)
* py3k fix in distributed LSI (spacecowboy, #433)
* Windows fix for (#428)
* fix compatibility for scipy 0.16.0 (#415)
0.12.1, 20/07/2015
* improvements to testing, switch to Travis CI containers
* support for loading old word2vec models (<=0.11.1) in 0.12+ (Gordon Mohr, #405)
* various bug fixes to word2vec, doc2vec (Gordon Mohr, #393, #386, #404)
* TextSummatization support for very short texts (Federico Barrios, #390)
* support for word2vec[['word1', 'word2'...]] convenience API calls (Satish Palaniappan, #395)
* MatrixSimilarity supports indexing generator corpora (single pass)
0.12.0, 06/07/2015
* complete API, performance, memory overhaul of doc2vec (Gordon Mohr, #356, #373, #380, #384)
- fast infer_vector(); optional memory-mapped doc vectors; memory savings with int doc IDs
- 'dbow_words' for combined DBOW & word skip-gram training; new 'dm_concat' mode
- multithreading & negative-sampling optimizations (also benefitting word2vec)
- API NOTE: doc vectors must now be accessed/compared through model's 'docvecs' field
(eg: "model.docvecs['my_ID']" or "model.docvecs.most_similar('my_ID')")
* new "text summarization" module (PR #324: Federico Lopez, Federico Barrios)
* new matutils.argsort with partial sort
- performance speedups to all similarity queries (word2vec, Similarity classes...)
* word2vec can compute likelihood scores for classification (Mat Addy, #358)
* word2vec supports "encoding" parameter when loading from C format, for non-utf8 models
* more memory-efficient word2vec training (#385)
* fixes to Python3 compatibility (Pavel Kalaidin #330, S-Eugene #369)
* enhancements to save/load format (Liang Bo Wang #363, Gordon Mohr #356)
- pickle defaults to protocol=2 for better py3 compatibility
* fixes and improvements to wiki parsing (Lukas Elmer #357, Excellent5 #333)
* fix to phrases scoring (Ikuya Yamada, #353)
* speed up of phrases generation (Dave Challis, #349)
* changes to multipass LDA training (Christopher Corley, #298)
* various doc improvements and fixes (Matti Lyra #331, Hongjoo Lee #334)
* fixes and improvements to LDA (Christopher Corley #323)
0.11.0 = 0.11.1 = 0.11.1-1, 10/04/2015
* added "topic ranking" to sort topics by coherence in LdaModel (jtmcmc, #311)
* new fast ShardedCorpus out-of-core corpus (Jan Hajic jr., #284)
* utils.smart_open now uses the smart_open package (#316)
* new wrapper for LDA in Vowpal Wabbit (Dave Challis, #304)
* improvements to the DtmModel wrapper (Yang Han, #272, #277)
* move wrappers for external modeling programs into a submodule (Christopher Corley, #295)
* allow transparent compression of NumPy files in save/load (Christopher Corley, #248)
* save/load methods now accept file handles, in addition to file names (macks22, #292)
* fixes to LdaMulticore on Windows (Feng Mai, #305)
* lots of small fixes & py3k compatibility improvements (Chyi-Kwei Yau, Daniel Nouri, Timothy Emerick, Juarez Bochi, Christopher Corley, Chirag Nagpal, Jan Hajic jr., Flávio Codeço Coelho)
* re-released as 0.11.1 and 0.11.1-1 because of a packaging bug
0.10.3, 17/11/2014
* added streamed phrases = collocation detection (Miguel Cabrera, #258)
* added param for multiple word2vec epochs (sebastienj, #243)
* added doc2vec (=paragraph2vec = extension of word2vec) model (Timothy Emerick, #231)
* initialize word2vec deterministically, for increased experiment reproducibility (KCzar, #240)
* all indexed corpora now allow full Python slicing syntax (Christopher Corley, #246)
* update distributed code for new Pyro4 API and py3k (Michael Brooks, Marco Bonzanini, #255, #249)
* fixes to six module version (Lars Buitinck, #259)
* fixes to (Maxim Avanov and Christopher Corley, #260, #251)
* ...and lots of minor fixes & updates all around
0.10.2, 18/09/2014
* new parallelized, LdaMulticore implementation (Jan Zikes, #232)
* Dynamic Topic Models (DTM) wrapper (Arttii, #205)
* word2vec compiled from bundled C file at install time: no more pyximport (#233)
* standardize show_/print_topics in LdaMallet (Benjamin Bray, #223)
* add new word2vec multiplicative objective (3CosMul) of Levy & Goldberg (Gordon Mohr, #224)
* preserve case in MALLET wrapper (mcburton, #222)
* support for matrix-valued topic/word prior eta in LdaModel (mjwillson, #208)
* py3k fix to SparseCorpus (Andreas Madsen, #234)
* fix to LowCorpus when switching dictionaries (Christopher Corley, #237)
0.10.1, 22/07/2014
* word2vec: new n_similarity method for comparing two sets of words (François Scharffe, #219)
* make LDA print/show topics parameters consistent with LSI (Bram Vandekerckhove, #201)
* add option for efficient word2vec subsampling (Gordon Mohr, #206)
* fix length calculation for corpora on empty files (Christopher Corley, #209)
* improve file cleanup of unit tests (Christopher Corley)
* more unit tests
* unicode now stored everywhere in gensim internally; accepted input stays either utf8 or unicode
* various fixes to the py3k ported code
* allow any dict-like input in Dictionary.from_corpus (Andreas Madsen)
* error checking improvements to the MALLET wrapper
* ignore non-articles during wiki parsig
* utils.lemmatize now (optionally) ignores stopwords
0.10.0 (aka "PY3K port"), 04/06/2014
* full Python 3 support (targeting 3.3+, #196)
* all internal methods now expect & store unicode, instead of utf8
* new optimized word2vec functionality: negative sampling, cbow (sebastien-j, #162)
* allow by-frequency sort in Dictionary.save_as_text (Renaud Richardet, #192)
* add topic printing to HDP model (Tiepes, #190)
* new gensim_addons package = optional install-time Cython compilations (Björn Esser, #197)
* added py3.3 and 3.4 to Travis CI tests
* fix a cbow word2vec bug (Liang-Chi Hsieh)
0.9.1, 12/04/2014
* MmCorpus fix for Windows
* LdaMallet support for printing/showing topics
* fix LdaMallet bug when user specified a file prefix (Victor, #184)
* fix LdaMallet output when input is single vector (Suvir)
* added LdaMallet unit tests
* more py3k fixes (Lars Buitinck)
* change order of LDA topic printing (Fayimora Femi-Balogun, #188)
0.9.0, 16/03/2014
* save/load automatically single out large arrays + allow mmap
* allow .gz/.bz2 corpus filenames => transparently (de)compressed I/O
* CBOW model for word2vec (Sébastien Jean, #176)
* new API for storing corpus metadata (Joseph Chang, #169)
* new LdaMallet class = train LDA using wrapped Mallet
* new MalletCorpus class for corpora in Mallet format (Christopher Corley, #179)
* better Wikipedia article parsing (Joseph Chang, #170)
* word2vec load_word2vec_format uses less memory (Yves Raimond, #164)
* load/store vocabulary files for word2vec C format (Yves Raimond, #172)
* HDP estimation on new documents (Elliot Kulakow, #153)
* store labels in SvmLight corpus (Ritesh, #152)
* fix word2vec binary load on Windows (Stephanus van Schalkwyk)
* replace numpy.svd with scipy.svd for more stability (Sven Döring, #159)
* parametrize LDA constructor (Christopher Corley, #174)
* steps toward py3k compatibility (Lars Buitinck, #154)
0.8.9, 26/12/2013
* use travis-ci for continuous integration
* auto-optimize LDA asymmetric prior (Ben Trahan)
* update for new word2vec binary format (Daren Race)
* doc rendering fix (Dan Foreman-Mackey)
* better LDA perplexity logging
* fix Pyro thread leak in distributed algos (Brian Feeny)
* optimizations in word2vec (Bryan Rink)
* allow compressed input in LineSentence corpus (Eric Moyer)
* upgrade ez_setup, doc improvements, minor fixes etc.
0.8.8 (aka "word2vec release"), 03/11/2013
* python3 port by Parikshit Samant:
* massive optimizations to word2vec (cython, BLAS, multithreading): ~20x-300x speedup
* new word2vec functionality (thx to Ghassen Hamrouni, PR #124)
* new CSV corpus class (thx to Zygmunt Zając)
* corpus serialization checks to prevent overwriting (by Ian Langmore, PR #125)
* add context manager support for older Python<=2.6 for gzip and bz2
* added unittests for word2vec
0.8.7, 18/09/2013
* initial version of word2vec, a neural network deep learning algo
* make distributed gensim compatible with the new Pyro
* allow merging dictionaries (by Florent Chandelier)
* new design for the gensim website!
* speed up handling of corner cases when returning top-n most similar
* make Random Projections compatible with new scipy (andrewjOc360, PR #110)
* allow "light" (faster) word lemmatization (by Karsten Jeschkies)
* save/load directly from bzip2 files (by Luis Pedro Coelho, PR #101)
* Blei corpus now tries harder to find its vocabulary file (by Luis Pedro Coelho, PR #100)
* sparse vector elements can now be a list (was: only a 2-tuple)
* simple_preprocess now optionally deaccents letters (ř/š/ú=>r/s/u etc.)
* better serialization of numpy corpora
* print_topics() returns the topics, in addition to printing/logging
* fixes for more robust Windows multiprocessing
* lots of small fixes, data checks and documentation updates
0.8.6, 15/09/2012
* added HashDictionary (by Homer Strong)
* support for adding target classes in SVMlight format (by Corrado Monti)
* fixed problems with global lemmatizer object when running in parallel on Windows
* parallelization of Wikipedia processing + added script version that lemmatizes the input documents
* added class method to initialize Dictionary from an existing corpus (by Marko Burjek)
0.8.5, 22/07/2012
* improved performance of sharding (similarity queries)
* better Wikipedia parsing (thx to Alejandro Weinstein and Lars Buitinck)
* faster Porter stemmer (thx to Lars Buitinck)
* several minor fixes (in HDP model thx to Greg Ver Steeg)
* improvements to documentation
0.8.4, 09/03/2012
* better support for Pandas series input (thx to JT Bates)
* a new corpus format: UCI bag-of-words (thx to Jonathan Esterhazy)
* a new model, non-parametric bayes: HDP (thx to Jonathan Esterhazy; based on Chong Wang's code)
* improved support for new scipy versions (thx to Skipper Seabold)
* lemmatizer support for wikipedia parsing (via the `pattern` python package)
* extended the lemmatizer for multi-core processing, to improve its performance
0.8.3, 02/12/2011
* fixed Similarity sharding bug (issue #65, thx to Paul Rudin)
* improved LDA code (clarity & memory footprint)
* optimized efficiency of Similarity sharding
0.8.2, 31/10/2011
* improved gensim landing page
* improved accuracy of SVD (Latent Semantic Analysis) (thx to Mark Tygert)
* changed interpretation of LDA topics: github issue #57
* took out similarity server code introduced in 0.8.1 (will become a separate project)
* started using `tox` for testing
* + several smaller fixes and optimizations
0.8.1, 10/10/2011
* transactional similarity server: see docs/simserver.html
* website moved from university hosting to
* much improved speed of lsi[corpus] transformation:
* accuracy tests of incremental svd: test/ and
* further improvements to memory-efficiency of LDA and LSA
* improved wiki preprocessing (thx to Luca de Alfaro)
* model.print_topics() debug fncs now support std output, in addition to logging (thx to Homer Strong)
* several smaller fixes and improvements
0.8.0 (Armageddon), 28/06/2011
* changed all variable and function names to comply with PEP8 (numTopics->num_topics): BREAKS BACKWARD COMPATIBILITY!
* added support for similarity querying more documents at once (index[query_documents] in addition to index[query_document]; much faster)
* rewrote Similarity so that it is more efficient and scalable (using disk-based mmap'ed shards)
* simplified directory structure (src/gensim/ is now only gensim/)
* several small fixes and optimizations
0.7.8, 26/03/2011
* added `corpora.IndexedCorpus`, a base class for corpus serializers (thx to Dieter Plaetinck). This allows corpus formats that inherit from it (MmCorpus, SvmLightCorpus, BleiCorpus etc.) to retrieve individual documents by their id in O(1), e.g. `corpus[14]` returns document #14.
* merged new code from the team (`corpora.textcorpus`, `models.logentropy_model`, lots of unit tests etc.)
* fixed a bug in `lda[bow]` transformation (was returning gamma distribution instead of theta). LDA model generation was not affected, only transforming new vectors.
* several small fixes and documentation updates
0.7.7, 13/02/2011
* new LDA implementation after Hoffman et al.: Online Learning for Latent Dirichlet Allocation
* distributed LDA
* updated LDA docs (wiki experiments, distributed tutorial)
* matrixmarket header now uses capital 'M's: MatrixMarket. (André Lynum reported than Matlab has trouble processing the lowercase version)
* moved code to github
* started gensim Google group
0.7.6, 10/01/2011
* added workaround for a bug in numpy: pickling a fortran-order array (e.g. LSA model) and then loading it back and using it results in segfault (thx to Brian Merrel)
* bundled a new version of old failed with Python2.6 when setuptools were missing (thx to Alan Salmoni).
0.7.5, 03/11/2010
* further optimization to LSA; this is the version used in my NIPS workshop paper
* got rid of SVDLIBC dependency (one-pass LSA now uses stochastic algo for base-base decompositions)
* sped up Latent Dirichlet ~10x (through scipy.weave, optional)
* finally, distributed LDA! scales almost linearly, but no tutorial yet. see the tutorial on distributed LSI, everything's completely analogous.
* several minor fixes and improvements; one nasty bug fixed (lsi[corpus] didn't work; thx to Danilo Spinelli)
* added stochastic SVD decomposition (faster than the current one-pass LSI algo, but needs two passes over the input corpus)
* published gensim on
* added workaround for a numpy bug where SVD sometimes fails to converge for no good reason
* changed content of gensims's PyPi title page
* completed HTML tutorial on distributed LSA
* fixed a bug in LSA that occurred when the number of features was smaller than the number of topics (thx to Richard Berendsen)
* optimized vocabulary generation in gensim.corpora.dictionary (faster and less memory-intense)
* MmCorpus accepts compressed input (file-like objects such as GzipFile, BZ2File; to save disk space)
* changed sparse solver to SVDLIBC (sparsesvd on PyPi) for large document chunks
* added distributed LSA, updated tutorials (still experimental though)
* several minor bug fixes
* added option for online LSI training (yay!). the transformation can now be
used after any amount of training, and training can be continued at any time
with more data.
* optimized the tf-idf transformation, so that it is a strictly one-pass algorithm in all cases (thx to Brian Merrell).
* fixed Windows-specific bug in handling binary files (thx to Sutee Sudprasert)
* fixed 1-based feature counting bug in SVMlight format (thx to Richard Berendsen)
* added 'Topic :: Text Processing :: Linguistic' to gensim's pypi classifiers
* change of sphinx documentation css and layout
* finished all tutorials, stable version
* tutorial on transformations
* added Random Projections (aka Random Indexing), as another transformation model.
* several DML-CZ specific updates
* updated documentation
* further memory optimizations in SVD (LSI)
* added missing test files to
* documentation changes
* added gensim reference to Wikipedia articles (SVD, LSI, LDA, TFIDF, ...)
* finally, a tutorial!
* similarity queries got their own package
* pdf documentation
* removed dependency on python2.5 (theoretically, gensim now runs on 2.6 and 2.7 as well).
* support for ``python test``
* fixing package metadata
* documentation clean-up
* First version
Something went wrong with that request. Please try again.