From 6afb3ff9bcaa4f5faef05a4223c324852b2e4c22 Mon Sep 17 00:00:00 2001
From: Paul Andrey <paul.andrey@hotmail.fr>
Date: Mon, 7 Jun 2021 17:52:16 +0200
Subject: [PATCH 01/13] Implemented `reduced_windows` argument for Word2Vec.

Co-Authored-By: Mathis Demay <mathis.demay.md@gmail.com>
---
 gensim/models/word2vec.py             | 43 +++++++++++++++++++--------
 gensim/models/word2vec_corpusfile.pyx | 24 +++++++++++----
 gensim/models/word2vec_inner.pyx      | 26 ++++++++++++----
 3 files changed, 68 insertions(+), 25 deletions(-)

diff --git a/gensim/models/word2vec.py b/gensim/models/word2vec.py
index 2593a373b0..e3c22dd95c 100755
--- a/gensim/models/word2vec.py
+++ b/gensim/models/word2vec.py
@@ -223,13 +223,13 @@
 
     def train_epoch_sg(
             model, corpus_file, offset, _cython_vocab, _cur_epoch, _expected_examples, _expected_words,
-            _work, _neu1, compute_loss,
+            _work, _neu1, compute_loss, reduced_windows,
         ):
         raise RuntimeError("Training with corpus_file argument is not supported")
 
     def train_epoch_cbow(
             model, corpus_file, offset, _cython_vocab, _cur_epoch, _expected_examples, _expected_words,
-            _work, _neu1, compute_loss,
+            _work, _neu1, compute_loss, reduced_windows,
         ):
         raise RuntimeError("Training with corpus_file argument is not supported")
 
@@ -240,7 +240,7 @@ def __init__(
             max_vocab_size=None, sample=1e-3, seed=1, workers=3, min_alpha=0.0001,
             sg=0, hs=0, negative=5, ns_exponent=0.75, cbow_mean=1, hashfxn=hash, epochs=5, null_word=0,
             trim_rule=None, sorted_vocab=1, batch_words=MAX_WORDS_IN_BATCH, compute_loss=False, callbacks=(),
-            comment=None, max_final_vocab=None,
+            comment=None, max_final_vocab=None, reduced_windows=True
         ):
         """Train, use and evaluate neural networks described in https://code.google.com/p/word2vec/.
 
@@ -345,6 +345,9 @@ def __init__(
             :meth:`~gensim.models.word2vec.Word2Vec.get_latest_training_loss`.
         callbacks : iterable of :class:`~gensim.models.callbacks.CallbackAny2Vec`, optional
             Sequence of callbacks to be executed at specific stages during training.
+        reduced_windows : bool, optional
+            If True, the window size is uniformly sampled from {1, `window`}
+            during training. Otherwise, it is fixed to `window`.
 
         Examples
         --------
@@ -377,6 +380,7 @@ def __init__(
         self.min_alpha = float(min_alpha)
 
         self.window = int(window)
+        self.reduced_windows = bool(reduced_windows)
         self.random = np.random.RandomState(seed)
 
         self.hs = int(hs)
@@ -419,7 +423,8 @@ def __init__(
             self.train(
                 corpus_iterable=corpus_iterable, corpus_file=corpus_file, total_examples=self.corpus_count,
                 total_words=self.corpus_total_words, epochs=self.epochs, start_alpha=self.alpha,
-                end_alpha=self.min_alpha, compute_loss=self.compute_loss, callbacks=callbacks)
+                end_alpha=self.min_alpha, compute_loss=self.compute_loss, callbacks=callbacks,
+                reduced_windows=self.reduced_windows)
         else:
             if trim_rule is not None:
                 logger.warning(
@@ -910,12 +915,14 @@ def _do_train_epoch(
         if self.sg:
             examples, tally, raw_tally = train_epoch_sg(
                 self, corpus_file, offset, cython_vocab, cur_epoch,
-                total_examples, total_words, work, neu1, self.compute_loss,
+                total_examples, total_words, work, neu1,
+                self.compute_loss, self.reduced_windows,
             )
         else:
             examples, tally, raw_tally = train_epoch_cbow(
                 self, corpus_file, offset, cython_vocab, cur_epoch,
-                total_examples, total_words, work, neu1, self.compute_loss,
+                total_examples, total_words, work, neu1,
+                self.compute_loss, self.reduced_windows,
             )
 
         return examples, tally, raw_tally
@@ -941,9 +948,15 @@ def _do_train_job(self, sentences, alpha, inits):
         work, neu1 = inits
         tally = 0
         if self.sg:
-            tally += train_batch_sg(self, sentences, alpha, work, self.compute_loss)
+            tally += train_batch_sg(
+                self, sentences, alpha, work,
+                self.compute_loss, self.reduced_windows,
+            )
         else:
-            tally += train_batch_cbow(self, sentences, alpha, work, neu1, self.compute_loss)
+            tally += train_batch_cbow(
+                self, sentences, alpha, work, neu1,
+                self.compute_loss, self.reduced_windows,
+            )
         return tally, self._raw_word_count(sentences)
 
     def _clear_post_train(self):
@@ -951,10 +964,10 @@ def _clear_post_train(self):
         self.wv.norms = None
 
     def train(
-            self, corpus_iterable=None, corpus_file=None, total_examples=None, total_words=None,
-            epochs=None, start_alpha=None, end_alpha=None, word_count=0,
-            queue_factor=2, report_delay=1.0, compute_loss=False, callbacks=(),
-            **kwargs,
+            self, corpus_iterable=None, corpus_file=None, total_examples=None,
+            total_words=None, epochs=None, start_alpha=None, end_alpha=None,
+            word_count=0, queue_factor=2, report_delay=1.0, compute_loss=False,
+            reduced_windows=True, callbacks=(), **kwargs,
         ):
         """Update the model's neural weights from a sequence of sentences.
 
@@ -1011,6 +1024,9 @@ def train(
         compute_loss: bool, optional
             If True, computes and stores loss value which can be retrieved using
             :meth:`~gensim.models.word2vec.Word2Vec.get_latest_training_loss`.
+        reduced_windows : bool, optional
+            If True, the window size is uniformly sampled from {1, `window`}
+            during training. Otherwise, it is fixed to `window`.
         callbacks : iterable of :class:`~gensim.models.callbacks.CallbackAny2Vec`, optional
             Sequence of callbacks to be executed at specific stages during training.
 
@@ -1030,6 +1046,7 @@ def train(
         self.alpha = start_alpha or self.alpha
         self.min_alpha = end_alpha or self.min_alpha
         self.epochs = epochs
+        self.reduced_windows = reduced_windows
 
         self._check_training_sanity(epochs=epochs, total_examples=total_examples, total_words=total_words)
         self._check_corpus_sanity(corpus_iterable=corpus_iterable, corpus_file=corpus_file, passes=epochs)
@@ -1039,7 +1056,7 @@ def train(
             msg=(
                 f"training model with {self.workers} workers on {len(self.wv)} vocabulary and "
                 f"{self.layer1_size} features, using sg={self.sg} hs={self.hs} sample={self.sample} "
-                f"negative={self.negative} window={self.window}"
+                f"negative={self.negative} window={self.window} reduced_windows={self.reduced_windows}"
             ),
         )
 
diff --git a/gensim/models/word2vec_corpusfile.pyx b/gensim/models/word2vec_corpusfile.pyx
index 19b9b8c165..a59ab721ea 100644
--- a/gensim/models/word2vec_corpusfile.pyx
+++ b/gensim/models/word2vec_corpusfile.pyx
@@ -186,7 +186,8 @@ cdef void prepare_c_structures_for_batch(
         vector[vector[string]] &sentences, int sample, int hs, int window, long long *total_words,
         int *effective_words, int *effective_sentences, unsigned long long *next_random,
         cvocab_t *vocab, int *sentence_idx, np.uint32_t *indexes, int *codelens,
-        np.uint8_t **codes, np.uint32_t **points, np.uint32_t *reduced_windows) nogil:
+        np.uint8_t **codes, np.uint32_t **points, np.uint32_t *reduced_windows,
+        bint do_reduced_windows) nogil:
     cdef VocabItem word
     cdef string token
     cdef vector[string] sent
@@ -225,7 +226,10 @@ cdef void prepare_c_structures_for_batch(
 
     # precompute "reduced window" offsets in a single randint() call
     for i in range(effective_words[0]):
-        reduced_windows[i] = random_int32(next_random) % window
+        if do_reduced_windows:
+            reduced_windows[i] = random_int32(next_random) % window
+        else:
+            reduced_windows[i] = window
 
 
 cdef REAL_t get_alpha(REAL_t alpha, REAL_t end_alpha, int cur_epoch, int num_epochs) nogil:
@@ -250,7 +254,7 @@ cdef REAL_t get_next_alpha(
 
 
 def train_epoch_sg(model, corpus_file, offset, _cython_vocab, _cur_epoch, _expected_examples, _expected_words, _work,
-                   _neu1, compute_loss):
+                   _neu1, compute_loss, reduced_windows):
     """Train Skipgram model for one epoch by training on an input stream. This function is used only in multistream mode.
 
     Called internally from :meth:`~gensim.models.word2vec.Word2Vec.train`.
@@ -269,6 +273,9 @@ def train_epoch_sg(model, corpus_file, offset, _cython_vocab, _cur_epoch, _expec
         Private working memory for each worker.
     compute_loss : bool
         Whether or not the training loss should be computed in this batch.
+    reduced_windows : bool
+        Whether or not the window size should be reduced based on random
+        uniform sampling.
 
     Returns
     -------
@@ -295,6 +302,7 @@ def train_epoch_sg(model, corpus_file, offset, _cython_vocab, _cur_epoch, _expec
     cdef long long total_sentences = 0
     cdef long long total_effective_words = 0, total_words = 0
     cdef int sent_idx, idx_start, idx_end
+    cdef bint do_reduced_windows = reduced_windows
 
     init_w2v_config(&c, model, _alpha, compute_loss, _work)
 
@@ -311,7 +319,7 @@ def train_epoch_sg(model, corpus_file, offset, _cython_vocab, _cur_epoch, _expec
             prepare_c_structures_for_batch(
                 sentences, c.sample, c.hs, c.window, &total_words, &effective_words, &effective_sentences,
                 &c.next_random, vocab.get_vocab_ptr(), c.sentence_idx, c.indexes,
-                c.codelens, c.codes, c.points, c.reduced_windows)
+                c.codelens, c.codes, c.points, c.reduced_windows, do_reduced_windows)
 
             for sent_idx in range(effective_sentences):
                 idx_start = c.sentence_idx[sent_idx]
@@ -350,7 +358,7 @@ def train_epoch_sg(model, corpus_file, offset, _cython_vocab, _cur_epoch, _expec
 
 
 def train_epoch_cbow(model, corpus_file, offset, _cython_vocab, _cur_epoch, _expected_examples, _expected_words, _work,
-                     _neu1, compute_loss):
+                     _neu1, compute_loss, reduced_windows):
     """Train CBOW model for one epoch by training on an input stream. This function is used only in multistream mode.
 
     Called internally from :meth:`~gensim.models.word2vec.Word2Vec.train`.
@@ -369,6 +377,9 @@ def train_epoch_cbow(model, corpus_file, offset, _cython_vocab, _cur_epoch, _exp
         Private working memory for each worker.
     compute_loss : bool
         Whether or not the training loss should be computed in this batch.
+    reduced_windows : bool
+        Whether or not the window size should be reduced based on random
+        uniform sampling.
 
     Returns
     -------
@@ -395,6 +406,7 @@ def train_epoch_cbow(model, corpus_file, offset, _cython_vocab, _cur_epoch, _exp
     cdef long long total_sentences = 0
     cdef long long total_effective_words = 0, total_words = 0
     cdef int sent_idx, idx_start, idx_end
+    cdef bint do_reduced_windows = reduced_windows
 
     init_w2v_config(&c, model, _alpha, compute_loss, _work, _neu1)
 
@@ -411,7 +423,7 @@ def train_epoch_cbow(model, corpus_file, offset, _cython_vocab, _cur_epoch, _exp
             prepare_c_structures_for_batch(
                 sentences, c.sample, c.hs, c.window, &total_words, &effective_words,
                 &effective_sentences, &c.next_random, vocab.get_vocab_ptr(), c.sentence_idx,
-                c.indexes, c.codelens, c.codes, c.points, c.reduced_windows)
+                c.indexes, c.codelens, c.codes, c.points, c.reduced_windows, do_reduced_windows)
 
             for sent_idx in range(effective_sentences):
                 idx_start = c.sentence_idx[sent_idx]
diff --git a/gensim/models/word2vec_inner.pyx b/gensim/models/word2vec_inner.pyx
index 9dc47fb040..64525094aa 100755
--- a/gensim/models/word2vec_inner.pyx
+++ b/gensim/models/word2vec_inner.pyx
@@ -71,10 +71,10 @@ cdef void our_saxpy_noblas(const int *N, const float *alpha, const float *X, con
     for i from 0 <= i < N[0] by 1:
         Y[i * (incY[0])] = (alpha[0]) * X[i * (incX[0])] + Y[i * (incY[0])]
 
-cdef long long _mul(const np.uint32_t a, const int b) nogil: 
+cdef long long _mul(const np.uint32_t a, const int b) nogil:
     """Safe multiplication of ints with explict typecasting"""
     return <long long>a * <long long>b
-        
+
 cdef void w2v_fast_sentence_sg_hs(
     const np.uint32_t *word_point, const np.uint8_t *word_code, const int codelen,
     REAL_t *syn0, REAL_t *syn1, const int size,
@@ -502,7 +502,7 @@ cdef init_w2v_config(Word2VecConfig *c, model, alpha, compute_loss, _work, _neu1
         c[0].neu1 = <REAL_t *>np.PyArray_DATA(_neu1)
 
 
-def train_batch_sg(model, sentences, alpha, _work, compute_loss):
+def train_batch_sg(model, sentences, alpha, _work, compute_loss, reduced_windows):
     """Update skip-gram model by training on a batch of sentences.
 
     Called internally from :meth:`~gensim.models.word2vec.Word2Vec.train`.
@@ -519,6 +519,9 @@ def train_batch_sg(model, sentences, alpha, _work, compute_loss):
         Private working memory for each worker.
     compute_loss : bool
         Whether or not the training loss should be computed in this batch.
+    reduced_windows : bool
+        Whether or not the window size should be reduced based on random
+        uniform sampling.
 
     Returns
     -------
@@ -570,7 +573,11 @@ def train_batch_sg(model, sentences, alpha, _work, compute_loss):
             break  # TODO: log warning, tally overflow?
 
     # precompute "reduced window" offsets in a single randint() call
-    for i, item in enumerate(model.random.randint(0, c.window, effective_words)):
+    if reduced_windows:
+        window_size = model.random.randint(0, c.window, effective_words)
+    else:
+        window_size = [0] * effective_words
+    for i, item in enumerate(window_size):
         c.reduced_windows[i] = item
 
     # release GIL & train on all sentences
@@ -597,7 +604,7 @@ def train_batch_sg(model, sentences, alpha, _work, compute_loss):
     return effective_words
 
 
-def train_batch_cbow(model, sentences, alpha, _work, _neu1, compute_loss):
+def train_batch_cbow(model, sentences, alpha, _work, _neu1, compute_loss, reduced_windows):
     """Update CBOW model by training on a batch of sentences.
 
     Called internally from :meth:`~gensim.models.word2vec.Word2Vec.train`.
@@ -616,6 +623,9 @@ def train_batch_cbow(model, sentences, alpha, _work, _neu1, compute_loss):
         Private working memory for each worker.
     compute_loss : bool
         Whether or not the training loss should be computed in this batch.
+    reduced_windows : bool
+        Whether or not the window size should be reduced based on random
+        uniform sampling.
 
     Returns
     -------
@@ -666,7 +676,11 @@ def train_batch_cbow(model, sentences, alpha, _work, _neu1, compute_loss):
             break  # TODO: log warning, tally overflow?
 
     # precompute "reduced window" offsets in a single randint() call
-    for i, item in enumerate(model.random.randint(0, c.window, effective_words)):
+    if reduced_windows:
+        window_size = model.random.randint(0, c.window, effective_words)
+    else:
+        window_size = [0] * effective_words
+    for i, item in enumerate(window_size):
         c.reduced_windows[i] = item
 
     # release GIL & train on all sentences

From 9bdb5a1ecb3e42a5cd789473a34a7413b923e139 Mon Sep 17 00:00:00 2001
From: Paul Andrey <paul.andrey@hotmail.fr>
Date: Thu, 10 Jun 2021 16:59:29 +0200
Subject: [PATCH 02/13] Improve the way `reduced_windows` is passed around and
 used.

---
 gensim/models/word2vec.py             | 31 ++++++++++++++++-----------
 gensim/models/word2vec_corpusfile.pyx | 18 ++++++----------
 gensim/models/word2vec_inner.pyx      | 28 +++++++++---------------
 3 files changed, 34 insertions(+), 43 deletions(-)

diff --git a/gensim/models/word2vec.py b/gensim/models/word2vec.py
index e3c22dd95c..53a02d80fe 100755
--- a/gensim/models/word2vec.py
+++ b/gensim/models/word2vec.py
@@ -223,13 +223,13 @@
 
     def train_epoch_sg(
             model, corpus_file, offset, _cython_vocab, _cur_epoch, _expected_examples, _expected_words,
-            _work, _neu1, compute_loss, reduced_windows,
+            _work, _neu1, compute_loss,
         ):
         raise RuntimeError("Training with corpus_file argument is not supported")
 
     def train_epoch_cbow(
             model, corpus_file, offset, _cython_vocab, _cur_epoch, _expected_examples, _expected_words,
-            _work, _neu1, compute_loss, reduced_windows,
+            _work, _neu1, compute_loss,
         ):
         raise RuntimeError("Training with corpus_file argument is not supported")
 
@@ -240,7 +240,7 @@ def __init__(
             max_vocab_size=None, sample=1e-3, seed=1, workers=3, min_alpha=0.0001,
             sg=0, hs=0, negative=5, ns_exponent=0.75, cbow_mean=1, hashfxn=hash, epochs=5, null_word=0,
             trim_rule=None, sorted_vocab=1, batch_words=MAX_WORDS_IN_BATCH, compute_loss=False, callbacks=(),
-            comment=None, max_final_vocab=None, reduced_windows=True
+            comment=None, max_final_vocab=None, reduced_windows=True,
         ):
         """Train, use and evaluate neural networks described in https://code.google.com/p/word2vec/.
 
@@ -346,8 +346,10 @@ def __init__(
         callbacks : iterable of :class:`~gensim.models.callbacks.CallbackAny2Vec`, optional
             Sequence of callbacks to be executed at specific stages during training.
         reduced_windows : bool, optional
-            If True, the window size is uniformly sampled from {1, `window`}
-            during training. Otherwise, it is fixed to `window`.
+            If True, the effective window size is uniformly sampled from  [1, `window`]
+            for each target word during training, to match the original word2vec algorithm's
+            approximate weighting of context words by distance. Otherwise, the effective
+            window size is always fixed to `window` words to either side.
 
         Examples
         --------
@@ -916,13 +918,13 @@ def _do_train_epoch(
             examples, tally, raw_tally = train_epoch_sg(
                 self, corpus_file, offset, cython_vocab, cur_epoch,
                 total_examples, total_words, work, neu1,
-                self.compute_loss, self.reduced_windows,
+                self.compute_loss,
             )
         else:
             examples, tally, raw_tally = train_epoch_cbow(
                 self, corpus_file, offset, cython_vocab, cur_epoch,
                 total_examples, total_words, work, neu1,
-                self.compute_loss, self.reduced_windows,
+                self.compute_loss,
             )
 
         return examples, tally, raw_tally
@@ -950,12 +952,12 @@ def _do_train_job(self, sentences, alpha, inits):
         if self.sg:
             tally += train_batch_sg(
                 self, sentences, alpha, work,
-                self.compute_loss, self.reduced_windows,
+                self.compute_loss,
             )
         else:
             tally += train_batch_cbow(
                 self, sentences, alpha, work, neu1,
-                self.compute_loss, self.reduced_windows,
+                self.compute_loss,
             )
         return tally, self._raw_word_count(sentences)
 
@@ -967,7 +969,7 @@ def train(
             self, corpus_iterable=None, corpus_file=None, total_examples=None,
             total_words=None, epochs=None, start_alpha=None, end_alpha=None,
             word_count=0, queue_factor=2, report_delay=1.0, compute_loss=False,
-            reduced_windows=True, callbacks=(), **kwargs,
+            reduced_windows=None, callbacks=(), **kwargs,
         ):
         """Update the model's neural weights from a sequence of sentences.
 
@@ -1025,8 +1027,10 @@ def train(
             If True, computes and stores loss value which can be retrieved using
             :meth:`~gensim.models.word2vec.Word2Vec.get_latest_training_loss`.
         reduced_windows : bool, optional
-            If True, the window size is uniformly sampled from {1, `window`}
-            during training. Otherwise, it is fixed to `window`.
+            If True, the effective window size is uniformly sampled from  [1, `window`]
+            for each target word during training, to match the original word2vec algorithm's
+            approximate weighting of context words by distance. Otherwise, the effective
+            window size is always fixed to `window` words to either side.
         callbacks : iterable of :class:`~gensim.models.callbacks.CallbackAny2Vec`, optional
             Sequence of callbacks to be executed at specific stages during training.
 
@@ -1046,7 +1050,8 @@ def train(
         self.alpha = start_alpha or self.alpha
         self.min_alpha = end_alpha or self.min_alpha
         self.epochs = epochs
-        self.reduced_windows = reduced_windows
+        if reduced_windows is not None:
+            self.reduced_windows = bool(reduced_windows)
 
         self._check_training_sanity(epochs=epochs, total_examples=total_examples, total_words=total_words)
         self._check_corpus_sanity(corpus_iterable=corpus_iterable, corpus_file=corpus_file, passes=epochs)
diff --git a/gensim/models/word2vec_corpusfile.pyx b/gensim/models/word2vec_corpusfile.pyx
index a59ab721ea..2af398e58b 100644
--- a/gensim/models/word2vec_corpusfile.pyx
+++ b/gensim/models/word2vec_corpusfile.pyx
@@ -187,7 +187,7 @@ cdef void prepare_c_structures_for_batch(
         int *effective_words, int *effective_sentences, unsigned long long *next_random,
         cvocab_t *vocab, int *sentence_idx, np.uint32_t *indexes, int *codelens,
         np.uint8_t **codes, np.uint32_t **points, np.uint32_t *reduced_windows,
-        bint do_reduced_windows) nogil:
+        int do_reduced_windows) nogil:
     cdef VocabItem word
     cdef string token
     cdef vector[string] sent
@@ -229,7 +229,7 @@ cdef void prepare_c_structures_for_batch(
         if do_reduced_windows:
             reduced_windows[i] = random_int32(next_random) % window
         else:
-            reduced_windows[i] = window
+            reduced_windows[i] = 0
 
 
 cdef REAL_t get_alpha(REAL_t alpha, REAL_t end_alpha, int cur_epoch, int num_epochs) nogil:
@@ -254,7 +254,7 @@ cdef REAL_t get_next_alpha(
 
 
 def train_epoch_sg(model, corpus_file, offset, _cython_vocab, _cur_epoch, _expected_examples, _expected_words, _work,
-                   _neu1, compute_loss, reduced_windows):
+                   _neu1, compute_loss,):
     """Train Skipgram model for one epoch by training on an input stream. This function is used only in multistream mode.
 
     Called internally from :meth:`~gensim.models.word2vec.Word2Vec.train`.
@@ -273,9 +273,6 @@ def train_epoch_sg(model, corpus_file, offset, _cython_vocab, _cur_epoch, _expec
         Private working memory for each worker.
     compute_loss : bool
         Whether or not the training loss should be computed in this batch.
-    reduced_windows : bool
-        Whether or not the window size should be reduced based on random
-        uniform sampling.
 
     Returns
     -------
@@ -302,7 +299,7 @@ def train_epoch_sg(model, corpus_file, offset, _cython_vocab, _cur_epoch, _expec
     cdef long long total_sentences = 0
     cdef long long total_effective_words = 0, total_words = 0
     cdef int sent_idx, idx_start, idx_end
-    cdef bint do_reduced_windows = reduced_windows
+    cdef int do_reduced_windows = int(model.reduced_windows)
 
     init_w2v_config(&c, model, _alpha, compute_loss, _work)
 
@@ -358,7 +355,7 @@ def train_epoch_sg(model, corpus_file, offset, _cython_vocab, _cur_epoch, _expec
 
 
 def train_epoch_cbow(model, corpus_file, offset, _cython_vocab, _cur_epoch, _expected_examples, _expected_words, _work,
-                     _neu1, compute_loss, reduced_windows):
+                     _neu1, compute_loss,):
     """Train CBOW model for one epoch by training on an input stream. This function is used only in multistream mode.
 
     Called internally from :meth:`~gensim.models.word2vec.Word2Vec.train`.
@@ -377,9 +374,6 @@ def train_epoch_cbow(model, corpus_file, offset, _cython_vocab, _cur_epoch, _exp
         Private working memory for each worker.
     compute_loss : bool
         Whether or not the training loss should be computed in this batch.
-    reduced_windows : bool
-        Whether or not the window size should be reduced based on random
-        uniform sampling.
 
     Returns
     -------
@@ -406,7 +400,7 @@ def train_epoch_cbow(model, corpus_file, offset, _cython_vocab, _cur_epoch, _exp
     cdef long long total_sentences = 0
     cdef long long total_effective_words = 0, total_words = 0
     cdef int sent_idx, idx_start, idx_end
-    cdef bint do_reduced_windows = reduced_windows
+    cdef int do_reduced_windows = int(model.reduced_windows)
 
     init_w2v_config(&c, model, _alpha, compute_loss, _work, _neu1)
 
diff --git a/gensim/models/word2vec_inner.pyx b/gensim/models/word2vec_inner.pyx
index 64525094aa..7389f70054 100755
--- a/gensim/models/word2vec_inner.pyx
+++ b/gensim/models/word2vec_inner.pyx
@@ -502,7 +502,7 @@ cdef init_w2v_config(Word2VecConfig *c, model, alpha, compute_loss, _work, _neu1
         c[0].neu1 = <REAL_t *>np.PyArray_DATA(_neu1)
 
 
-def train_batch_sg(model, sentences, alpha, _work, compute_loss, reduced_windows):
+def train_batch_sg(model, sentences, alpha, _work, compute_loss):
     """Update skip-gram model by training on a batch of sentences.
 
     Called internally from :meth:`~gensim.models.word2vec.Word2Vec.train`.
@@ -519,9 +519,6 @@ def train_batch_sg(model, sentences, alpha, _work, compute_loss, reduced_windows
         Private working memory for each worker.
     compute_loss : bool
         Whether or not the training loss should be computed in this batch.
-    reduced_windows : bool
-        Whether or not the window size should be reduced based on random
-        uniform sampling.
 
     Returns
     -------
@@ -573,12 +570,11 @@ def train_batch_sg(model, sentences, alpha, _work, compute_loss, reduced_windows
             break  # TODO: log warning, tally overflow?
 
     # precompute "reduced window" offsets in a single randint() call
-    if reduced_windows:
-        window_size = model.random.randint(0, c.window, effective_words)
+    if model.reduced_windows:
+        for i, item in enumerate(model.random.randint(0, c.window, effective_words)):
+            c.reduced_windows[i] = item
     else:
-        window_size = [0] * effective_words
-    for i, item in enumerate(window_size):
-        c.reduced_windows[i] = item
+        c.reduced_windows[:] = int(0)
 
     # release GIL & train on all sentences
     with nogil:
@@ -604,7 +600,7 @@ def train_batch_sg(model, sentences, alpha, _work, compute_loss, reduced_windows
     return effective_words
 
 
-def train_batch_cbow(model, sentences, alpha, _work, _neu1, compute_loss, reduced_windows):
+def train_batch_cbow(model, sentences, alpha, _work, _neu1, compute_loss):
     """Update CBOW model by training on a batch of sentences.
 
     Called internally from :meth:`~gensim.models.word2vec.Word2Vec.train`.
@@ -623,9 +619,6 @@ def train_batch_cbow(model, sentences, alpha, _work, _neu1, compute_loss, reduce
         Private working memory for each worker.
     compute_loss : bool
         Whether or not the training loss should be computed in this batch.
-    reduced_windows : bool
-        Whether or not the window size should be reduced based on random
-        uniform sampling.
 
     Returns
     -------
@@ -676,12 +669,11 @@ def train_batch_cbow(model, sentences, alpha, _work, _neu1, compute_loss, reduce
             break  # TODO: log warning, tally overflow?
 
     # precompute "reduced window" offsets in a single randint() call
-    if reduced_windows:
-        window_size = model.random.randint(0, c.window, effective_words)
+    if model.reduced_windows:
+        for i, item in enumerate(model.random.randint(0, c.window, effective_words)):
+            c.reduced_windows[i] = item
     else:
-        window_size = [0] * effective_words
-    for i, item in enumerate(window_size):
-        c.reduced_windows[i] = item
+        c.reduced_windows[:] = int(0)
 
     # release GIL & train on all sentences
     with nogil:

From 77698cd502c268a41fed3db5ba3833f95d8576f7 Mon Sep 17 00:00:00 2001
From: Paul Andrey <paul.andrey@hotmail.fr>
Date: Thu, 10 Jun 2021 17:01:45 +0200
Subject: [PATCH 03/13] Renamed `reduced_windows` to `shrink_windows`.

---
 gensim/models/word2vec.py             | 18 +++++++++---------
 gensim/models/word2vec_corpusfile.pyx | 12 ++++++------
 gensim/models/word2vec_inner.pyx      |  4 ++--
 3 files changed, 17 insertions(+), 17 deletions(-)

diff --git a/gensim/models/word2vec.py b/gensim/models/word2vec.py
index 53a02d80fe..aa245c369a 100755
--- a/gensim/models/word2vec.py
+++ b/gensim/models/word2vec.py
@@ -240,7 +240,7 @@ def __init__(
             max_vocab_size=None, sample=1e-3, seed=1, workers=3, min_alpha=0.0001,
             sg=0, hs=0, negative=5, ns_exponent=0.75, cbow_mean=1, hashfxn=hash, epochs=5, null_word=0,
             trim_rule=None, sorted_vocab=1, batch_words=MAX_WORDS_IN_BATCH, compute_loss=False, callbacks=(),
-            comment=None, max_final_vocab=None, reduced_windows=True,
+            comment=None, max_final_vocab=None, shrink_windows=True,
         ):
         """Train, use and evaluate neural networks described in https://code.google.com/p/word2vec/.
 
@@ -345,7 +345,7 @@ def __init__(
             :meth:`~gensim.models.word2vec.Word2Vec.get_latest_training_loss`.
         callbacks : iterable of :class:`~gensim.models.callbacks.CallbackAny2Vec`, optional
             Sequence of callbacks to be executed at specific stages during training.
-        reduced_windows : bool, optional
+        shrink_windows : bool, optional
             If True, the effective window size is uniformly sampled from  [1, `window`]
             for each target word during training, to match the original word2vec algorithm's
             approximate weighting of context words by distance. Otherwise, the effective
@@ -382,7 +382,7 @@ def __init__(
         self.min_alpha = float(min_alpha)
 
         self.window = int(window)
-        self.reduced_windows = bool(reduced_windows)
+        self.shrink_windows = bool(shrink_windows)
         self.random = np.random.RandomState(seed)
 
         self.hs = int(hs)
@@ -426,7 +426,7 @@ def __init__(
                 corpus_iterable=corpus_iterable, corpus_file=corpus_file, total_examples=self.corpus_count,
                 total_words=self.corpus_total_words, epochs=self.epochs, start_alpha=self.alpha,
                 end_alpha=self.min_alpha, compute_loss=self.compute_loss, callbacks=callbacks,
-                reduced_windows=self.reduced_windows)
+                shrink_windows=self.shrink_windows)
         else:
             if trim_rule is not None:
                 logger.warning(
@@ -969,7 +969,7 @@ def train(
             self, corpus_iterable=None, corpus_file=None, total_examples=None,
             total_words=None, epochs=None, start_alpha=None, end_alpha=None,
             word_count=0, queue_factor=2, report_delay=1.0, compute_loss=False,
-            reduced_windows=None, callbacks=(), **kwargs,
+            shrink_windows=None, callbacks=(), **kwargs,
         ):
         """Update the model's neural weights from a sequence of sentences.
 
@@ -1026,7 +1026,7 @@ def train(
         compute_loss: bool, optional
             If True, computes and stores loss value which can be retrieved using
             :meth:`~gensim.models.word2vec.Word2Vec.get_latest_training_loss`.
-        reduced_windows : bool, optional
+        shrink_windows : bool, optional
             If True, the effective window size is uniformly sampled from  [1, `window`]
             for each target word during training, to match the original word2vec algorithm's
             approximate weighting of context words by distance. Otherwise, the effective
@@ -1050,8 +1050,8 @@ def train(
         self.alpha = start_alpha or self.alpha
         self.min_alpha = end_alpha or self.min_alpha
         self.epochs = epochs
-        if reduced_windows is not None:
-            self.reduced_windows = bool(reduced_windows)
+        if shrink_windows is not None:
+            self.shrink_windows = bool(shrink_windows)
 
         self._check_training_sanity(epochs=epochs, total_examples=total_examples, total_words=total_words)
         self._check_corpus_sanity(corpus_iterable=corpus_iterable, corpus_file=corpus_file, passes=epochs)
@@ -1061,7 +1061,7 @@ def train(
             msg=(
                 f"training model with {self.workers} workers on {len(self.wv)} vocabulary and "
                 f"{self.layer1_size} features, using sg={self.sg} hs={self.hs} sample={self.sample} "
-                f"negative={self.negative} window={self.window} reduced_windows={self.reduced_windows}"
+                f"negative={self.negative} window={self.window} shrink_windows={self.shrink_windows}"
             ),
         )
 
diff --git a/gensim/models/word2vec_corpusfile.pyx b/gensim/models/word2vec_corpusfile.pyx
index 2af398e58b..da94e78ec1 100644
--- a/gensim/models/word2vec_corpusfile.pyx
+++ b/gensim/models/word2vec_corpusfile.pyx
@@ -187,7 +187,7 @@ cdef void prepare_c_structures_for_batch(
         int *effective_words, int *effective_sentences, unsigned long long *next_random,
         cvocab_t *vocab, int *sentence_idx, np.uint32_t *indexes, int *codelens,
         np.uint8_t **codes, np.uint32_t **points, np.uint32_t *reduced_windows,
-        int do_reduced_windows) nogil:
+        int shrink_windows) nogil:
     cdef VocabItem word
     cdef string token
     cdef vector[string] sent
@@ -226,7 +226,7 @@ cdef void prepare_c_structures_for_batch(
 
     # precompute "reduced window" offsets in a single randint() call
     for i in range(effective_words[0]):
-        if do_reduced_windows:
+        if shrink_windows:
             reduced_windows[i] = random_int32(next_random) % window
         else:
             reduced_windows[i] = 0
@@ -299,7 +299,7 @@ def train_epoch_sg(model, corpus_file, offset, _cython_vocab, _cur_epoch, _expec
     cdef long long total_sentences = 0
     cdef long long total_effective_words = 0, total_words = 0
     cdef int sent_idx, idx_start, idx_end
-    cdef int do_reduced_windows = int(model.reduced_windows)
+    cdef int shrink_windows = int(model.shrink_windows)
 
     init_w2v_config(&c, model, _alpha, compute_loss, _work)
 
@@ -316,7 +316,7 @@ def train_epoch_sg(model, corpus_file, offset, _cython_vocab, _cur_epoch, _expec
             prepare_c_structures_for_batch(
                 sentences, c.sample, c.hs, c.window, &total_words, &effective_words, &effective_sentences,
                 &c.next_random, vocab.get_vocab_ptr(), c.sentence_idx, c.indexes,
-                c.codelens, c.codes, c.points, c.reduced_windows, do_reduced_windows)
+                c.codelens, c.codes, c.points, c.reduced_windows, shrink_windows)
 
             for sent_idx in range(effective_sentences):
                 idx_start = c.sentence_idx[sent_idx]
@@ -400,7 +400,7 @@ def train_epoch_cbow(model, corpus_file, offset, _cython_vocab, _cur_epoch, _exp
     cdef long long total_sentences = 0
     cdef long long total_effective_words = 0, total_words = 0
     cdef int sent_idx, idx_start, idx_end
-    cdef int do_reduced_windows = int(model.reduced_windows)
+    cdef int shrink_windows = int(model.shrink_windows)
 
     init_w2v_config(&c, model, _alpha, compute_loss, _work, _neu1)
 
@@ -417,7 +417,7 @@ def train_epoch_cbow(model, corpus_file, offset, _cython_vocab, _cur_epoch, _exp
             prepare_c_structures_for_batch(
                 sentences, c.sample, c.hs, c.window, &total_words, &effective_words,
                 &effective_sentences, &c.next_random, vocab.get_vocab_ptr(), c.sentence_idx,
-                c.indexes, c.codelens, c.codes, c.points, c.reduced_windows, do_reduced_windows)
+                c.indexes, c.codelens, c.codes, c.points, c.reduced_windows, shrink_windows)
 
             for sent_idx in range(effective_sentences):
                 idx_start = c.sentence_idx[sent_idx]
diff --git a/gensim/models/word2vec_inner.pyx b/gensim/models/word2vec_inner.pyx
index 7389f70054..cd3b5feefa 100755
--- a/gensim/models/word2vec_inner.pyx
+++ b/gensim/models/word2vec_inner.pyx
@@ -570,7 +570,7 @@ def train_batch_sg(model, sentences, alpha, _work, compute_loss):
             break  # TODO: log warning, tally overflow?
 
     # precompute "reduced window" offsets in a single randint() call
-    if model.reduced_windows:
+    if model.shrink_windows:
         for i, item in enumerate(model.random.randint(0, c.window, effective_words)):
             c.reduced_windows[i] = item
     else:
@@ -669,7 +669,7 @@ def train_batch_cbow(model, sentences, alpha, _work, _neu1, compute_loss):
             break  # TODO: log warning, tally overflow?
 
     # precompute "reduced window" offsets in a single randint() call
-    if model.reduced_windows:
+    if model.shrink_windows:
         for i, item in enumerate(model.random.randint(0, c.window, effective_words)):
             c.reduced_windows[i] = item
     else:

From 11a464ffdae1fbc8e29f44400722d5cf626e9edb Mon Sep 17 00:00:00 2001
From: Paul Andrey <paul.andrey@hotmail.fr>
Date: Fri, 11 Jun 2021 18:24:04 +0200
Subject: [PATCH 04/13] Removed `shrink_windows` argument from
 `Word2Vec.train`.

---
 gensim/models/word2vec.py | 7 ++-----
 1 file changed, 2 insertions(+), 5 deletions(-)

diff --git a/gensim/models/word2vec.py b/gensim/models/word2vec.py
index aa245c369a..8b8160fa12 100755
--- a/gensim/models/word2vec.py
+++ b/gensim/models/word2vec.py
@@ -425,8 +425,7 @@ def __init__(
             self.train(
                 corpus_iterable=corpus_iterable, corpus_file=corpus_file, total_examples=self.corpus_count,
                 total_words=self.corpus_total_words, epochs=self.epochs, start_alpha=self.alpha,
-                end_alpha=self.min_alpha, compute_loss=self.compute_loss, callbacks=callbacks,
-                shrink_windows=self.shrink_windows)
+                end_alpha=self.min_alpha, compute_loss=self.compute_loss, callbacks=callbacks)
         else:
             if trim_rule is not None:
                 logger.warning(
@@ -969,7 +968,7 @@ def train(
             self, corpus_iterable=None, corpus_file=None, total_examples=None,
             total_words=None, epochs=None, start_alpha=None, end_alpha=None,
             word_count=0, queue_factor=2, report_delay=1.0, compute_loss=False,
-            shrink_windows=None, callbacks=(), **kwargs,
+            callbacks=(), **kwargs,
         ):
         """Update the model's neural weights from a sequence of sentences.
 
@@ -1050,8 +1049,6 @@ def train(
         self.alpha = start_alpha or self.alpha
         self.min_alpha = end_alpha or self.min_alpha
         self.epochs = epochs
-        if shrink_windows is not None:
-            self.shrink_windows = bool(shrink_windows)
 
         self._check_training_sanity(epochs=epochs, total_examples=total_examples, total_words=total_words)
         self._check_corpus_sanity(corpus_iterable=corpus_iterable, corpus_file=corpus_file, passes=epochs)

From a65751889960e7a76ebe3471cf2268d9b289a5b7 Mon Sep 17 00:00:00 2001
From: Paul Andrey <paul.andrey@hotmail.fr>
Date: Tue, 15 Jun 2021 22:47:24 +0200
Subject: [PATCH 05/13] Aesthetic fix.

---
 gensim/models/word2vec_corpusfile.pyx | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/gensim/models/word2vec_corpusfile.pyx b/gensim/models/word2vec_corpusfile.pyx
index da94e78ec1..13b69d30da 100644
--- a/gensim/models/word2vec_corpusfile.pyx
+++ b/gensim/models/word2vec_corpusfile.pyx
@@ -187,7 +187,8 @@ cdef void prepare_c_structures_for_batch(
         int *effective_words, int *effective_sentences, unsigned long long *next_random,
         cvocab_t *vocab, int *sentence_idx, np.uint32_t *indexes, int *codelens,
         np.uint8_t **codes, np.uint32_t **points, np.uint32_t *reduced_windows,
-        int shrink_windows) nogil:
+        int shrink_windows,
+    ) nogil:
     cdef VocabItem word
     cdef string token
     cdef vector[string] sent

From ff0d30e3ae79f52a3e5fff3b7f32d83ff5a8b6e8 Mon Sep 17 00:00:00 2001
From: Paul Andrey <paul.andrey@hotmail.fr>
Date: Tue, 15 Jun 2021 22:48:04 +0200
Subject: [PATCH 06/13] Fixed old word2vec models' reloading.

---
 gensim/models/word2vec.py | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/gensim/models/word2vec.py b/gensim/models/word2vec.py
index 8b8160fa12..6c9a826a59 100755
--- a/gensim/models/word2vec.py
+++ b/gensim/models/word2vec.py
@@ -1989,6 +1989,8 @@ def _load_specials(self, *args, **kwargs):
                 self.syn1 = self.syn1
                 del self.syn1
             del self.trainables
+        if not hasattr(self, 'shrink_windows'):
+            self.shrink_windows = True
 
     def get_latest_training_loss(self):
         """Get current value of the training loss.

From f98eb43c4aea38de32640b42b24d385bd73d5e12 Mon Sep 17 00:00:00 2001
From: Paul Andrey <paul.andrey@hotmail.fr>
Date: Tue, 15 Jun 2021 22:55:22 +0200
Subject: [PATCH 07/13] Fixed undue docstring.

---
 gensim/models/word2vec.py | 5 -----
 1 file changed, 5 deletions(-)

diff --git a/gensim/models/word2vec.py b/gensim/models/word2vec.py
index 6c9a826a59..91817c2b15 100755
--- a/gensim/models/word2vec.py
+++ b/gensim/models/word2vec.py
@@ -1025,11 +1025,6 @@ def train(
         compute_loss: bool, optional
             If True, computes and stores loss value which can be retrieved using
             :meth:`~gensim.models.word2vec.Word2Vec.get_latest_training_loss`.
-        shrink_windows : bool, optional
-            If True, the effective window size is uniformly sampled from  [1, `window`]
-            for each target word during training, to match the original word2vec algorithm's
-            approximate weighting of context words by distance. Otherwise, the effective
-            window size is always fixed to `window` words to either side.
         callbacks : iterable of :class:`~gensim.models.callbacks.CallbackAny2Vec`, optional
             Sequence of callbacks to be executed at specific stages during training.
 

From 3b27f1c3cb6e3ee20d75b7d2834816f60b49cac5 Mon Sep 17 00:00:00 2001
From: Paul Andrey <paul.andrey@hotmail.fr>
Date: Tue, 15 Jun 2021 22:55:49 +0200
Subject: [PATCH 08/13] Added `shrink_windows` argument to Doc2Vec.

---
 gensim/models/doc2vec.py             |  8 +++++++-
 gensim/models/doc2vec_corpusfile.pyx | 17 ++++++++++++-----
 gensim/models/doc2vec_inner.pyx      | 16 +++++++++++-----
 3 files changed, 30 insertions(+), 11 deletions(-)

diff --git a/gensim/models/doc2vec.py b/gensim/models/doc2vec.py
index 23f7d9fc7e..072bb8e233 100644
--- a/gensim/models/doc2vec.py
+++ b/gensim/models/doc2vec.py
@@ -158,7 +158,7 @@ def count(self, new_val):
 class Doc2Vec(Word2Vec):
     def __init__(self, documents=None, corpus_file=None, vector_size=100, dm_mean=None, dm=1, dbow_words=0, dm_concat=0,
                  dm_tag_count=1, dv=None, dv_mapfile=None, comment=None, trim_rule=None, callbacks=(),
-                 window=5, epochs=10, **kwargs):
+                 window=5, epochs=10, shrink_windows=True, **kwargs):
         """Class for training, using and evaluating neural networks described in
         `Distributed Representations of Sentences and Documents <http://arxiv.org/abs/1405.4053v2>`_.
 
@@ -248,6 +248,11 @@ def __init__(self, documents=None, corpus_file=None, vector_size=100, dm_mean=No
 
         callbacks : :obj: `list` of :obj: `~gensim.models.callbacks.CallbackAny2Vec`, optional
             List of callbacks that need to be executed/run at specific stages during training.
+        shrink_windows : bool, optional
+            If True, the effective window size is uniformly sampled from  [1, `window`]
+            for each target word during training, to match the original word2vec algorithm's
+            approximate weighting of context words by distance. Otherwise, the effective
+            window size is always fixed to `window` words to either side.
 
         Some important internal attributes are the following:
 
@@ -294,6 +299,7 @@ def __init__(self, documents=None, corpus_file=None, vector_size=100, dm_mean=No
             callbacks=callbacks,
             window=window,
             epochs=epochs,
+            shrink_windows=shrink_windows,
             **kwargs,
         )
 
diff --git a/gensim/models/doc2vec_corpusfile.pyx b/gensim/models/doc2vec_corpusfile.pyx
index 40bf20bdd3..a7173ca026 100644
--- a/gensim/models/doc2vec_corpusfile.pyx
+++ b/gensim/models/doc2vec_corpusfile.pyx
@@ -59,7 +59,7 @@ cdef void prepare_c_structures_for_batch(
         int *effective_words, unsigned long long *next_random, cvocab_t *vocab,
         np.uint32_t *indexes, int *codelens, np.uint8_t **codes, np.uint32_t **points,
         np.uint32_t *reduced_windows, int *document_len, int train_words,
-        int docvecs_count, int doc_tag,
+        int docvecs_count, int doc_tag, int shrink_windows,
     ) nogil:
     cdef VocabItem predict_word
     cdef string token
@@ -88,7 +88,10 @@ cdef void prepare_c_structures_for_batch(
 
     if train_words and reduced_windows != NULL:
         for i in range(document_len[0]):
-            reduced_windows[i] = random_int32(next_random) % window
+            if shrink_windows:
+                reduced_windows[i] = random_int32(next_random) % window
+            else:
+                reduced_windows[i] = 0
 
     if doc_tag < docvecs_count:
         effective_words[0] += 1
@@ -160,6 +163,7 @@ def d2v_train_epoch_dbow(
     cdef long long total_documents = 0
     cdef long long total_effective_words = 0, total_words = 0
     cdef int sent_idx, idx_start, idx_end
+    cdef int shrink_windows = int(model.shrink_windows)
 
     cdef vector[string] doc_words
     cdef long long _doc_tag = start_doctag
@@ -183,7 +187,7 @@ def d2v_train_epoch_dbow(
             prepare_c_structures_for_batch(
                 doc_words, c.sample, c.hs, c.window, &total_words, &effective_words,
                 &c.next_random, vocab.get_vocab_ptr(), c.indexes, c.codelens,  c.codes, c.points,
-                c.reduced_windows, &document_len, c.train_words, c.docvecs_count, _doc_tag)
+                c.reduced_windows, &document_len, c.train_words, c.docvecs_count, _doc_tag, shrink_windows)
 
             for i in range(document_len):
                 if c.train_words:  # simultaneous skip-gram wordvec-training
@@ -300,6 +304,7 @@ def d2v_train_epoch_dm(
     cdef long long total_effective_words = 0, total_words = 0
     cdef int sent_idx, idx_start, idx_end
     cdef REAL_t count, inv_count = 1.0
+    cdef int shrink_windows = int(model.shrink_windows)
 
     cdef vector[string] doc_words
     cdef long long _doc_tag = start_doctag
@@ -323,7 +328,7 @@ def d2v_train_epoch_dm(
             prepare_c_structures_for_batch(
                 doc_words, c.sample, c.hs, c.window, &total_words, &effective_words, &c.next_random,
                 vocab.get_vocab_ptr(), c.indexes, c.codelens, c.codes, c.points, c.reduced_windows,
-                &document_len, c.train_words, c.docvecs_count, _doc_tag)
+                &document_len, c.train_words, c.docvecs_count, _doc_tag, shrink_windows)
 
             for i in range(document_len):
                 j = i - c.window + c.reduced_windows[i]
@@ -453,6 +458,7 @@ def d2v_train_epoch_dm_concat(
     cdef long long total_documents = 0
     cdef long long total_effective_words = 0, total_words = 0
     cdef int sent_idx, idx_start, idx_end
+    cdef int shrink_windows = int(model.shrink_windows)
 
     cdef vector[string] doc_words
     cdef long long _doc_tag = start_doctag
@@ -490,7 +496,8 @@ def d2v_train_epoch_dm_concat(
             prepare_c_structures_for_batch(
                 doc_words, c.sample, c.hs, c.window, &total_words, &effective_words,
                 &c.next_random, vocab.get_vocab_ptr(), c.indexes, c.codelens, c.codes,
-                c.points, NULL, &document_len, c.train_words, c.docvecs_count, _doc_tag)
+                c.points, NULL, &document_len, c.train_words, c.docvecs_count, _doc_tag,
+                shrink_windows)
 
             for i in range(document_len):
                 j = i - c.window      # negative OK: will pad with null word
diff --git a/gensim/models/doc2vec_inner.pyx b/gensim/models/doc2vec_inner.pyx
index 23ede53c90..ce1e214c3a 100644
--- a/gensim/models/doc2vec_inner.pyx
+++ b/gensim/models/doc2vec_inner.pyx
@@ -364,9 +364,12 @@ def train_document_dbow(model, doc_words, doctag_indexes, alpha, work=None,
     c.document_len = i
 
     if c.train_words:
-        # single randint() call avoids a big thread-synchronization slowdown
-        for i, item in enumerate(model.random.randint(0, c.window, c.document_len)):
-            c.reduced_windows[i] = item
+        if model.shrink_windows:
+            # single randint() call avoids a big thread-synchronization slowdown
+            for i, item in enumerate(model.random.randint(0, c.window, c.document_len)):
+                c.reduced_windows[i] = item
+        else:
+            c.reduced_windows[:] = int(0)
 
     for i in range(c.doctag_len):
         c.doctag_indexes[i] = doctag_indexes[i]
@@ -497,8 +500,11 @@ def train_document_dm(model, doc_words, doctag_indexes, alpha, work=None, neu1=N
     c.document_len = i
 
     # single randint() call avoids a big thread-sync slowdown
-    for i, item in enumerate(model.random.randint(0, c.window, c.document_len)):
-        c.reduced_windows[i] = item
+    if model.shrink_windows:
+        for i, item in enumerate(model.random.randint(0, c.window, c.document_len)):
+            c.reduced_windows[i] = item
+    else:
+        c.reduced_windows[:] = int(0)
 
     for i in range(c.doctag_len):
         c.doctag_indexes[i] = doctag_indexes[i]

From 27a49d77dde635bc8c499eed157d70a76966234d Mon Sep 17 00:00:00 2001
From: Paul Andrey <paul.andrey@hotmail.fr>
Date: Tue, 15 Jun 2021 23:07:53 +0200
Subject: [PATCH 09/13] Added `shrink_windows` argument to FastText.

---
 gensim/models/fasttext.py             | 10 ++++++++--
 gensim/models/fasttext_corpusfile.pyx | 14 ++++++++++----
 gensim/models/fasttext_inner.pyx      |  7 +++++--
 3 files changed, 23 insertions(+), 8 deletions(-)

diff --git a/gensim/models/fasttext.py b/gensim/models/fasttext.py
index 61c093a3be..8d396400f2 100644
--- a/gensim/models/fasttext.py
+++ b/gensim/models/fasttext.py
@@ -276,7 +276,7 @@ def __init__(self, sentences=None, corpus_file=None, sg=0, hs=0, vector_size=100
                  max_vocab_size=None, word_ngrams=1, sample=1e-3, seed=1, workers=3, min_alpha=0.0001,
                  negative=5, ns_exponent=0.75, cbow_mean=1, hashfxn=hash, epochs=5, null_word=0, min_n=3, max_n=6,
                  sorted_vocab=1, bucket=2000000, trim_rule=None, batch_words=MAX_WORDS_IN_BATCH, callbacks=(),
-                 max_final_vocab=None):
+                 max_final_vocab=None, shrink_windows=True,):
         """Train, use and evaluate word representations learned using the method
         described in `Enriching Word Vectors with Subword Information <https://arxiv.org/abs/1607.04606>`_,
         aka FastText.
@@ -385,6 +385,11 @@ def __init__(self, sentences=None, corpus_file=None, sg=0, hs=0, vector_size=100
             ``min_count```.  If the specified ``min_count`` is more than the
             automatically calculated ``min_count``, the former will be used.
             Set to ``None`` if not required.
+        shrink_windows : bool, optional
+            If True, the effective window size is uniformly sampled from  [1, `window`]
+            for each target word during training, to match the original word2vec algorithm's
+            approximate weighting of context words by distance. Otherwise, the effective
+            window size is always fixed to `window` words to either side.
 
         Examples
         --------
@@ -432,7 +437,8 @@ def __init__(self, sentences=None, corpus_file=None, sg=0, hs=0, vector_size=100
             max_vocab_size=max_vocab_size, max_final_vocab=max_final_vocab,
             min_count=min_count, sample=sample, sorted_vocab=sorted_vocab,
             null_word=null_word, ns_exponent=ns_exponent, hashfxn=hashfxn,
-            seed=seed, hs=hs, negative=negative, cbow_mean=cbow_mean, min_alpha=min_alpha)
+            seed=seed, hs=hs, negative=negative, cbow_mean=cbow_mean,
+            min_alpha=min_alpha, shrink_windows=shrink_windows)
 
     def _init_post_load(self, hidden_output):
         num_vectors = len(self.wv.vectors)
diff --git a/gensim/models/fasttext_corpusfile.pyx b/gensim/models/fasttext_corpusfile.pyx
index e5ec611aa0..d2abb4c04f 100644
--- a/gensim/models/fasttext_corpusfile.pyx
+++ b/gensim/models/fasttext_corpusfile.pyx
@@ -46,7 +46,8 @@ cdef void prepare_c_structures_for_batch(
         vector[vector[string]] &sentences, int sample, int hs, int window, long long *total_words,
         int *effective_words, int *effective_sentences, unsigned long long *next_random, cvocab_t *vocab,
         int *sentence_idx, np.uint32_t *indexes, int *codelens, np.uint8_t **codes, np.uint32_t **points,
-        np.uint32_t *reduced_windows, int *subwords_idx_len, np.uint32_t **subwords_idx) nogil:
+        np.uint32_t *reduced_windows, int *subwords_idx_len, np.uint32_t **subwords_idx, int shrink_windows,
+    ) nogil:
     cdef VocabItem word
     cdef string token
     cdef vector[string] sent
@@ -89,7 +90,10 @@ cdef void prepare_c_structures_for_batch(
 
     # precompute "reduced window" offsets in a single randint() call
     for i in range(effective_words[0]):
-        reduced_windows[i] = random_int32(next_random) % window
+        if shrink_windows:
+            reduced_windows[i] = random_int32(next_random) % window
+        else:
+            reduced_windows[i] = 0
 
 
 def train_epoch_sg(
@@ -136,6 +140,7 @@ def train_epoch_sg(
     cdef long long total_sentences = 0
     cdef long long total_effective_words = 0, total_words = 0
     cdef int sent_idx, idx_start, idx_end
+    cdef int shrink_windows = int(model.shrink_windows)
 
     init_ft_config(&c, model, _alpha, _work, _l1)
 
@@ -153,7 +158,7 @@ def train_epoch_sg(
             prepare_c_structures_for_batch(
                 sentences, c.sample, c.hs, c.window, &total_words, &effective_words, &effective_sentences,
                 &c.next_random, vocab.get_vocab_ptr(), c.sentence_idx, c.indexes, c.codelens,
-                c.codes, c.points, c.reduced_windows, c.subwords_idx_len, c.subwords_idx)
+                c.codes, c.points, c.reduced_windows, c.subwords_idx_len, c.subwords_idx, shrink_windows)
 
             for sent_idx in range(effective_sentences):
                 idx_start = c.sentence_idx[sent_idx]
@@ -226,6 +231,7 @@ def train_epoch_cbow(model, corpus_file, offset, _cython_vocab, _cur_epoch, _exp
     cdef long long total_sentences = 0
     cdef long long total_effective_words = 0, total_words = 0
     cdef int sent_idx, idx_start, idx_end
+    cdef int shrink_windows = int(model.shrink_windows)
 
     init_ft_config(&c, model, _alpha, _work, _neu1)
 
@@ -243,7 +249,7 @@ def train_epoch_cbow(model, corpus_file, offset, _cython_vocab, _cur_epoch, _exp
             prepare_c_structures_for_batch(
                 sentences, c.sample, c.hs, c.window, &total_words, &effective_words, &effective_sentences,
                 &c.next_random, vocab.get_vocab_ptr(), c.sentence_idx, c.indexes, c.codelens,
-                c.codes, c.points, c.reduced_windows, c.subwords_idx_len, c.subwords_idx)
+                c.codes, c.points, c.reduced_windows, c.subwords_idx_len, c.subwords_idx, shrink_windows)
 
             for sent_idx in range(effective_sentences):
                 idx_start = c.sentence_idx[sent_idx]
diff --git a/gensim/models/fasttext_inner.pyx b/gensim/models/fasttext_inner.pyx
index e71ed6f31d..7416ce7dd0 100644
--- a/gensim/models/fasttext_inner.pyx
+++ b/gensim/models/fasttext_inner.pyx
@@ -601,8 +601,11 @@ def train_batch_any(model, sentences, alpha, _work, _neu1):
     num_words, num_sentences = populate_ft_config(&c, model.wv, model.wv.buckets_word, sentences)
 
     # precompute "reduced window" offsets in a single randint() call
-    for i, randint in enumerate(model.random.randint(0, c.window, num_words)):
-        c.reduced_windows[i] = randint
+    if model.shrink_windows:
+        for i, randint in enumerate(model.random.randint(0, c.window, num_words)):
+            c.reduced_windows[i] = randint
+    else:
+        c.reduced_windows[:] = int(0)
 
     # release GIL & train on all sentences in the batch
     with nogil:

From 6fde7e89e438efbd36495fe6081510c8a7fdb78f Mon Sep 17 00:00:00 2001
From: Paul Andrey <paul.andrey@hotmail.fr>
Date: Thu, 17 Jun 2021 10:45:06 +0200
Subject: [PATCH 10/13] Fixed and optimized `shrink_windows` backend use.

* `c.reduced_windows[:] = 0` syntax is not supported;
  as a consequence, all zero-value assignments due to
  `shrink_windows=False` have been rewritten as `for`
  loops.
* Those changes have now been tested (cf. next commit).
* NOTE: another way to proceed could be to initialize
  the `reduced_windows` array with zeros as values; it
  would then only be altered when `shrink_windows=True`.
---
 gensim/models/doc2vec_corpusfile.pyx  | 7 ++++---
 gensim/models/doc2vec_inner.pyx       | 8 +++++---
 gensim/models/fasttext_corpusfile.pyx | 7 ++++---
 gensim/models/fasttext_inner.pyx      | 3 ++-
 gensim/models/word2vec_corpusfile.pyx | 7 ++++---
 gensim/models/word2vec_inner.pyx      | 6 ++++--
 6 files changed, 23 insertions(+), 15 deletions(-)

diff --git a/gensim/models/doc2vec_corpusfile.pyx b/gensim/models/doc2vec_corpusfile.pyx
index a7173ca026..9216d13bd4 100644
--- a/gensim/models/doc2vec_corpusfile.pyx
+++ b/gensim/models/doc2vec_corpusfile.pyx
@@ -87,10 +87,11 @@ cdef void prepare_c_structures_for_batch(
     document_len[0] = i
 
     if train_words and reduced_windows != NULL:
-        for i in range(document_len[0]):
-            if shrink_windows:
+        if shrink_windows:
+            for i in range(document_len[0]):
                 reduced_windows[i] = random_int32(next_random) % window
-            else:
+        else:
+            for i in range(document_len[0]):
                 reduced_windows[i] = 0
 
     if doc_tag < docvecs_count:
diff --git a/gensim/models/doc2vec_inner.pyx b/gensim/models/doc2vec_inner.pyx
index ce1e214c3a..1657c59787 100644
--- a/gensim/models/doc2vec_inner.pyx
+++ b/gensim/models/doc2vec_inner.pyx
@@ -364,12 +364,13 @@ def train_document_dbow(model, doc_words, doctag_indexes, alpha, work=None,
     c.document_len = i
 
     if c.train_words:
+        # single randint() call avoids a big thread-synchronization slowdown
         if model.shrink_windows:
-            # single randint() call avoids a big thread-synchronization slowdown
             for i, item in enumerate(model.random.randint(0, c.window, c.document_len)):
                 c.reduced_windows[i] = item
         else:
-            c.reduced_windows[:] = int(0)
+            for i in range(c.document_len):
+                c.reduced_windows[i] = 0
 
     for i in range(c.doctag_len):
         c.doctag_indexes[i] = doctag_indexes[i]
@@ -504,7 +505,8 @@ def train_document_dm(model, doc_words, doctag_indexes, alpha, work=None, neu1=N
         for i, item in enumerate(model.random.randint(0, c.window, c.document_len)):
             c.reduced_windows[i] = item
     else:
-        c.reduced_windows[:] = int(0)
+        for i in range(c.document_len):
+            c.reduced_windows[i] = 0
 
     for i in range(c.doctag_len):
         c.doctag_indexes[i] = doctag_indexes[i]
diff --git a/gensim/models/fasttext_corpusfile.pyx b/gensim/models/fasttext_corpusfile.pyx
index d2abb4c04f..5d275b42b6 100644
--- a/gensim/models/fasttext_corpusfile.pyx
+++ b/gensim/models/fasttext_corpusfile.pyx
@@ -89,10 +89,11 @@ cdef void prepare_c_structures_for_batch(
             break
 
     # precompute "reduced window" offsets in a single randint() call
-    for i in range(effective_words[0]):
-        if shrink_windows:
+    if shrink_windows:
+        for i in range(effective_words[0]):
             reduced_windows[i] = random_int32(next_random) % window
-        else:
+    else:
+        for i in range(effective_words[0]):
             reduced_windows[i] = 0
 
 
diff --git a/gensim/models/fasttext_inner.pyx b/gensim/models/fasttext_inner.pyx
index 7416ce7dd0..e27bd62feb 100644
--- a/gensim/models/fasttext_inner.pyx
+++ b/gensim/models/fasttext_inner.pyx
@@ -605,7 +605,8 @@ def train_batch_any(model, sentences, alpha, _work, _neu1):
         for i, randint in enumerate(model.random.randint(0, c.window, num_words)):
             c.reduced_windows[i] = randint
     else:
-        c.reduced_windows[:] = int(0)
+        for i in range(num_words):
+            c.reduced_windows[i] = 0
 
     # release GIL & train on all sentences in the batch
     with nogil:
diff --git a/gensim/models/word2vec_corpusfile.pyx b/gensim/models/word2vec_corpusfile.pyx
index 13b69d30da..5d7f5004e4 100644
--- a/gensim/models/word2vec_corpusfile.pyx
+++ b/gensim/models/word2vec_corpusfile.pyx
@@ -226,10 +226,11 @@ cdef void prepare_c_structures_for_batch(
             break  # TODO: log warning, tally overflow?
 
     # precompute "reduced window" offsets in a single randint() call
-    for i in range(effective_words[0]):
-        if shrink_windows:
+    if shrink_windows:
+        for i in range(effective_words[0]):
             reduced_windows[i] = random_int32(next_random) % window
-        else:
+    else:
+        for i in range(effective_words[0]):
             reduced_windows[i] = 0
 
 
diff --git a/gensim/models/word2vec_inner.pyx b/gensim/models/word2vec_inner.pyx
index a351dc04cd..ffdc908b5c 100755
--- a/gensim/models/word2vec_inner.pyx
+++ b/gensim/models/word2vec_inner.pyx
@@ -570,7 +570,8 @@ def train_batch_sg(model, sentences, alpha, _work, compute_loss):
         for i, item in enumerate(model.random.randint(0, c.window, effective_words)):
             c.reduced_windows[i] = item
     else:
-        c.reduced_windows[:] = int(0)
+        for i in range(effective_words):
+            c.reduced_windows[i] = 0
 
     # release GIL & train on all sentences
     with nogil:
@@ -669,7 +670,8 @@ def train_batch_cbow(model, sentences, alpha, _work, _neu1, compute_loss):
         for i, item in enumerate(model.random.randint(0, c.window, effective_words)):
             c.reduced_windows[i] = item
     else:
-        c.reduced_windows[:] = int(0)
+        for i in range(effective_words):
+            c.reduced_windows[i] = 0
 
     # release GIL & train on all sentences
     with nogil:

From 486b3f502e45b4bc06bf1a9196c9de41587478b1 Mon Sep 17 00:00:00 2001
From: Paul Andrey <paul.andrey@hotmail.fr>
Date: Thu, 17 Jun 2021 11:15:26 +0200
Subject: [PATCH 11/13] Added tests for `shrink_windows=False` in
 Word2Vec-based models.

---
 gensim/test/test_doc2vec.py  | 38 ++++++++++++++++++++++++++++++++++++
 gensim/test/test_fasttext.py | 28 ++++++++++++++++++--------
 gensim/test/test_word2vec.py | 32 ++++++++++++++++++++++++++++++
 3 files changed, 90 insertions(+), 8 deletions(-)

diff --git a/gensim/test/test_doc2vec.py b/gensim/test/test_doc2vec.py
index 60c9158744..ebb1af7cbe 100644
--- a/gensim/test/test_doc2vec.py
+++ b/gensim/test/test_doc2vec.py
@@ -589,6 +589,44 @@ def test_dmc_neg_fromfile(self):
             )
             self.model_sanity(model)
 
+    def test_dmm_fixedwindowsize(self):
+        """Test DMM doc2vec training with fixed window size."""
+        model = doc2vec.Doc2Vec(
+            list_corpus, vector_size=24,
+            dm=1, dm_mean=1, window=4, shrink_windows=False,
+            hs=0, negative=10, alpha=0.05, min_count=2, epochs=20
+        )
+        self.model_sanity(model)
+
+    def test_dmm_fixedwindowsize_fromfile(self):
+        """Test DMM doc2vec training with fixed window size, from file."""
+        with temporary_file(get_tmpfile('gensim_doc2vec.tst')) as corpus_file:
+            save_lee_corpus_as_line_sentence(corpus_file)
+            model = doc2vec.Doc2Vec(
+                corpus_file=corpus_file, vector_size=24,
+                dm=1, dm_mean=1, window=4, shrink_windows=False,
+                hs=0, negative=10, alpha=0.05, min_count=2, epochs=20
+            )
+            self.model_sanity(model)
+
+    def test_dbow_fixedwindowsize(self):
+        """Test DBOW doc2vec training with fixed window size."""
+        model = doc2vec.Doc2Vec(
+            list_corpus, vector_size=16, shrink_windows=False,
+            dm=0, hs=0, negative=5, min_count=2, epochs=20
+        )
+        self.model_sanity(model)
+
+    def test_dbow_fixedwindowsize_fromfile(self):
+        """Test DBOW doc2vec training with fixed window size, from file."""
+        with temporary_file(get_tmpfile('gensim_doc2vec.tst')) as corpus_file:
+            save_lee_corpus_as_line_sentence(corpus_file)
+            model = doc2vec.Doc2Vec(
+                corpus_file=corpus_file, vector_size=16, shrink_windows=False,
+                dm=0, hs=0, negative=5, min_count=2, epochs=20
+            )
+            self.model_sanity(model)
+
     def test_parallel(self):
         """Test doc2vec parallel training with more than default 3 threads."""
         # repeat the ~300 doc (~60000 word) Lee corpus to get 6000 docs (~1.2M words)
diff --git a/gensim/test/test_fasttext.py b/gensim/test/test_fasttext.py
index a91a9000a7..382ecdfb2c 100644
--- a/gensim/test/test_fasttext.py
+++ b/gensim/test/test_fasttext.py
@@ -397,12 +397,12 @@ def test_wm_distance(self):
         dist = self.test_model.wv.wmdistance(doc, oov_doc)
         self.assertNotEqual(float('inf'), dist)
 
-    def test_cbow_hs_training(self):
+    def test_cbow_hs_training(self, shrink_windows=True):
 
         model_gensim = FT_gensim(
             vector_size=48, sg=0, cbow_mean=1, alpha=0.05, window=5, hs=1, negative=0,
             min_count=5, epochs=10, batch_words=1000, word_ngrams=1, sample=1e-3, min_n=3, max_n=6,
-            sorted_vocab=1, workers=1, min_alpha=0.0, bucket=BUCKET)
+            sorted_vocab=1, workers=1, min_alpha=0.0, bucket=BUCKET, shrink_windows=shrink_windows)
 
         lee_data = LineSentence(datapath('lee_background.cor'))
         model_gensim.build_vocab(lee_data)
@@ -429,12 +429,12 @@ def test_cbow_hs_training(self):
             overlap_count, 2,
             "only %i overlap in expected %s & actual %s" % (overlap_count, expected_sims_words, sims_gensim_words))
 
-    def test_cbow_hs_training_fromfile(self):
+    def test_cbow_hs_training_fromfile(self, shrink_windows=True):
         with temporary_file('gensim_fasttext.tst') as corpus_file:
             model_gensim = FT_gensim(
                 vector_size=48, sg=0, cbow_mean=1, alpha=0.05, window=5, hs=1, negative=0,
                 min_count=5, epochs=10, batch_words=1000, word_ngrams=1, sample=1e-3, min_n=3, max_n=6,
-                sorted_vocab=1, workers=1, min_alpha=0.0, bucket=BUCKET * 4)
+                sorted_vocab=1, workers=1, min_alpha=0.0, bucket=BUCKET * 4, shrink_windows=shrink_windows)
 
             lee_data = LineSentence(datapath('lee_background.cor'))
             utils.save_as_line_sentence(lee_data, corpus_file)
@@ -465,12 +465,12 @@ def test_cbow_hs_training_fromfile(self):
                 overlap_count, 2,
                 "only %i overlap in expected %s & actual %s" % (overlap_count, expected_sims_words, sims_gensim_words))
 
-    def test_sg_hs_training(self):
+    def test_sg_hs_training(self, shrink_windows=True):
 
         model_gensim = FT_gensim(
             vector_size=48, sg=1, cbow_mean=1, alpha=0.025, window=5, hs=1, negative=0,
             min_count=5, epochs=10, batch_words=1000, word_ngrams=1, sample=1e-3, min_n=3, max_n=6,
-            sorted_vocab=1, workers=1, min_alpha=0.0, bucket=BUCKET)
+            sorted_vocab=1, workers=1, min_alpha=0.0, bucket=BUCKET, shrink_windows=shrink_windows)
 
         lee_data = LineSentence(datapath('lee_background.cor'))
         model_gensim.build_vocab(lee_data)
@@ -497,12 +497,12 @@ def test_sg_hs_training(self):
             overlap_count, 2,
             "only %i overlap in expected %s & actual %s" % (overlap_count, expected_sims_words, sims_gensim_words))
 
-    def test_sg_hs_training_fromfile(self):
+    def test_sg_hs_training_fromfile(self, shrink_windows=True):
         with temporary_file('gensim_fasttext.tst') as corpus_file:
             model_gensim = FT_gensim(
                 vector_size=48, sg=1, cbow_mean=1, alpha=0.025, window=5, hs=1, negative=0,
                 min_count=5, epochs=10, batch_words=1000, word_ngrams=1, sample=1e-3, min_n=3, max_n=6,
-                sorted_vocab=1, workers=1, min_alpha=0.0, bucket=BUCKET)
+                sorted_vocab=1, workers=1, min_alpha=0.0, bucket=BUCKET, shrink_windows=shrink_windows)
 
             lee_data = LineSentence(datapath('lee_background.cor'))
             utils.save_as_line_sentence(lee_data, corpus_file)
@@ -533,6 +533,18 @@ def test_sg_hs_training_fromfile(self):
                 overlap_count, 2,
                 "only %i overlap in expected %s & actual %s" % (overlap_count, expected_sims_words, sims_gensim_words))
 
+    def test_cbow_hs_training_fixedwindowsize(self):
+        self.test_cbow_hs_training(shrink_windows=False)
+
+    def test_cbow_hs_training_fixedwindowsize_fromfile(self):
+        self.test_cbow_hs_training_fromfile(shrink_windows=False)
+
+    def test_sg_hs_training_fixedwindowsize(self):
+        self.test_sg_hs_training(shrink_windows=False)
+
+    def test_sg_hs_training_fixedwindowsize_fromfile(self):
+        self.test_sg_hs_training_fromfile(shrink_windows=False)
+
     def test_cbow_neg_training(self):
 
         model_gensim = FT_gensim(
diff --git a/gensim/test/test_word2vec.py b/gensim/test/test_word2vec.py
index d46b2f3e37..e85cee0d5d 100644
--- a/gensim/test/test_word2vec.py
+++ b/gensim/test/test_word2vec.py
@@ -686,6 +686,38 @@ def test_cbow_neg_fromfile(self):
         )
         self.model_sanity(model, with_corpus_file=True)
 
+    def test_sg_fixedwindowsize(self):
+        """Test skipgram with fixed window size. Use NS."""
+        model = word2vec.Word2Vec(
+            sg=1, window=5, shrink_windows=False, hs=0,
+            negative=15, min_count=5, epochs=10, workers=2
+        )
+        self.model_sanity(model)
+
+    def test_sg_fixedwindowsize_fromfile(self):
+        """Test skipgram with fixed window size. Use HS and train from file."""
+        model = word2vec.Word2Vec(
+            sg=1, window=5, shrink_windows=False, hs=1,
+            negative=0, min_count=5, epochs=10, workers=2
+        )
+        self.model_sanity(model, with_corpus_file=True)
+
+    def test_cbow_fixedwindowsize(self, ranks=None):
+        """Test CBOW with fixed window size. Use HS."""
+        model = word2vec.Word2Vec(
+            sg=0, cbow_mean=1, alpha=0.1, window=5, shrink_windows=False,
+            hs=1, negative=0, min_count=5, epochs=10, workers=2
+        )
+        self.model_sanity(model, ranks=ranks)
+
+    def test_cbow_fixedwindowsize_fromfile(self):
+        """Test CBOW with fixed window size. Use NS and train from file."""
+        model = word2vec.Word2Vec(
+            sg=0, cbow_mean=1, alpha=0.1, window=5, shrink_windows=False,
+            hs=0, negative=15, min_count=5, epochs=10, workers=2
+        )
+        self.model_sanity(model, with_corpus_file=True)
+
     def test_cosmul(self):
         model = word2vec.Word2Vec(sentences, vector_size=2, min_count=1, hs=1, negative=0)
         sims = model.wv.most_similar_cosmul('graph', topn=10)

From f0af84ead2777b76e0898bec730e07babefcf0fc Mon Sep 17 00:00:00 2001
From: Paul Andrey <paul.andrey@hotmail.fr>
Date: Thu, 17 Jun 2021 11:16:36 +0200
Subject: [PATCH 12/13] Added docstring mentions of `shrink_window` being
 experimental.

---
 gensim/models/doc2vec.py  | 1 +
 gensim/models/fasttext.py | 1 +
 gensim/models/word2vec.py | 1 +
 3 files changed, 3 insertions(+)

diff --git a/gensim/models/doc2vec.py b/gensim/models/doc2vec.py
index 072bb8e233..e6fcb72f0e 100644
--- a/gensim/models/doc2vec.py
+++ b/gensim/models/doc2vec.py
@@ -249,6 +249,7 @@ def __init__(self, documents=None, corpus_file=None, vector_size=100, dm_mean=No
         callbacks : :obj: `list` of :obj: `~gensim.models.callbacks.CallbackAny2Vec`, optional
             List of callbacks that need to be executed/run at specific stages during training.
         shrink_windows : bool, optional
+            New in 4.1. Experimental.
             If True, the effective window size is uniformly sampled from  [1, `window`]
             for each target word during training, to match the original word2vec algorithm's
             approximate weighting of context words by distance. Otherwise, the effective
diff --git a/gensim/models/fasttext.py b/gensim/models/fasttext.py
index 8d396400f2..ba08a15a92 100644
--- a/gensim/models/fasttext.py
+++ b/gensim/models/fasttext.py
@@ -386,6 +386,7 @@ def __init__(self, sentences=None, corpus_file=None, sg=0, hs=0, vector_size=100
             automatically calculated ``min_count``, the former will be used.
             Set to ``None`` if not required.
         shrink_windows : bool, optional
+            New in 4.1. Experimental.
             If True, the effective window size is uniformly sampled from  [1, `window`]
             for each target word during training, to match the original word2vec algorithm's
             approximate weighting of context words by distance. Otherwise, the effective
diff --git a/gensim/models/word2vec.py b/gensim/models/word2vec.py
index 91817c2b15..50999acb3f 100755
--- a/gensim/models/word2vec.py
+++ b/gensim/models/word2vec.py
@@ -346,6 +346,7 @@ def __init__(
         callbacks : iterable of :class:`~gensim.models.callbacks.CallbackAny2Vec`, optional
             Sequence of callbacks to be executed at specific stages during training.
         shrink_windows : bool, optional
+            New in 4.1. Experimental.
             If True, the effective window size is uniformly sampled from  [1, `window`]
             for each target word during training, to match the original word2vec algorithm's
             approximate weighting of context words by distance. Otherwise, the effective

From 49cf50ecf010138a0573346e34e952f0c79dc1b0 Mon Sep 17 00:00:00 2001
From: Paul Andrey <paul.andrey@hotmail.fr>
Date: Tue, 22 Jun 2021 14:54:17 +0200
Subject: [PATCH 13/13] Rolled back some purely aesthetic changes.

---
 gensim/models/word2vec.py | 24 ++++++++----------------
 1 file changed, 8 insertions(+), 16 deletions(-)

diff --git a/gensim/models/word2vec.py b/gensim/models/word2vec.py
index 50999acb3f..265364890b 100755
--- a/gensim/models/word2vec.py
+++ b/gensim/models/word2vec.py
@@ -917,14 +917,12 @@ def _do_train_epoch(
         if self.sg:
             examples, tally, raw_tally = train_epoch_sg(
                 self, corpus_file, offset, cython_vocab, cur_epoch,
-                total_examples, total_words, work, neu1,
-                self.compute_loss,
+                total_examples, total_words, work, neu1, self.compute_loss
             )
         else:
             examples, tally, raw_tally = train_epoch_cbow(
                 self, corpus_file, offset, cython_vocab, cur_epoch,
-                total_examples, total_words, work, neu1,
-                self.compute_loss,
+                total_examples, total_words, work, neu1, self.compute_loss
             )
 
         return examples, tally, raw_tally
@@ -950,15 +948,9 @@ def _do_train_job(self, sentences, alpha, inits):
         work, neu1 = inits
         tally = 0
         if self.sg:
-            tally += train_batch_sg(
-                self, sentences, alpha, work,
-                self.compute_loss,
-            )
+            tally += train_batch_sg(self, sentences, alpha, work, self.compute_loss)
         else:
-            tally += train_batch_cbow(
-                self, sentences, alpha, work, neu1,
-                self.compute_loss,
-            )
+            tally += train_batch_cbow(self, sentences, alpha, work, neu1, self.compute_loss)
         return tally, self._raw_word_count(sentences)
 
     def _clear_post_train(self):
@@ -966,10 +958,10 @@ def _clear_post_train(self):
         self.wv.norms = None
 
     def train(
-            self, corpus_iterable=None, corpus_file=None, total_examples=None,
-            total_words=None, epochs=None, start_alpha=None, end_alpha=None,
-            word_count=0, queue_factor=2, report_delay=1.0, compute_loss=False,
-            callbacks=(), **kwargs,
+            self, corpus_iterable=None, corpus_file=None, total_examples=None, total_words=None,
+            epochs=None, start_alpha=None, end_alpha=None, word_count=0,
+            queue_factor=2, report_delay=1.0, compute_loss=False, callbacks=(),
+            **kwargs,
         ):
         """Update the model's neural weights from a sequence of sentences.