diff --git a/hw11_natural_language_processing/martink_word2vec/printouts b/hw11_natural_language_processing/martink_word2vec/printouts
new file mode 100644
index 0000000000000000000000000000000000000000..fb384a63c4aa361aad9f62fd09e39c3bc82b771a
--- /dev/null
+++ b/hw11_natural_language_processing/martink_word2vec/printouts
@@ -0,0 +1,244 @@
+bash-4.3$ python word2vec_basic.py
+Found and verified text8.zip
+Data size 17005207
+Most common words (+UNK) [['UNK', 418391], ('the', 1061396), ('of', 593677), ('and', 416629), ('one', 411764)]
+Sample data [5239, 3084, 12, 6, 195, 2, 3137, 46, 59, 156] ['anarchism', 'originated', 'as', 'a', 'term', 'of', 'abuse', 'first', 'used', 'against']
+3084 originated -> 12 as
+3084 originated -> 5239 anarchism
+12 as -> 3084 originated
+12 as -> 6 a
+6 a -> 195 term
+6 a -> 12 as
+195 term -> 2 of
+195 term -> 6 a
+Initialized
+Average loss at step 0 : 260.113525391
+Nearest to a: expired, colony, vassar, documentation, renamo, parades, uncompetitive, garbled,
+Nearest to years: rolf, sciences, arid, subtlety, benzene, prefecture, ventilated, tupac,
+Nearest to may: gabor, moussa, marathon, mccain, mondo, robbie, ordinations, koenigsegg,
+Nearest to new: reznor, monophosphate, bonnie, kornilov, subordinates, arcas, holiness, offices,
+Nearest to all: unfold, newton, campbell, rote, dong, kabbalist, newly, tice,
+Nearest to used: collision, contents, phonetic, besht, fondly, unguarded, millay, servicio,
+Nearest to called: landing, pi, acetylcholine, attorney, entrance, importer, crowe, position,
+Nearest to that: prohibit, habit, comedian, mythical, mir, financi, sweyn, mages,
+Nearest to three: reeves, wounding, blass, humans, myth, leukemia, ritually, manitoba,
+Nearest to the: consisting, gibb, oman, seleucids, semiarid, promote, vectorborne, macs,
+Nearest to when: interference, carnivores, wco, eine, suppress, continual, linton, teachings,
+Nearest to its: vidzeme, aoc, missile, grapple, caterpillars, accessed, aaas, boyle,
+Nearest to use: rigorously, ashoka, weizenbaum, shedding, white, thuringiensis, gettysburg, expectation,
+Nearest to he: dont, hutchins, whit, withhold, reina, push, anatolian, illustrious,
+Nearest to six: equilateral, viscosity, self, zed, epa, ecuadorian, pap, gisela,
+Nearest to s: mendeleev, vulgaris, pack, zangger, blind, respectful, attica, grenades,
+Average loss at step 2000 : 113.902854166
+Average loss at step 4000 : 52.8573114848
+Average loss at step 6000 : 33.0519484727
+Average loss at step 8000 : 24.1580560323
+Average loss at step 10000 : 17.9564257891
+Nearest to a: the, this, and, UNK, influence, scrimmage, victoriae, analogue,
+Nearest to years: sciences, amsterdam, nine, victoriae, six, arid, mathbf, reginae,
+Nearest to may: meeting, developers, psychology, plot, spread, amyotrophic, alpina, koenigsegg,
+Nearest to new: and, gb, circles, go, offices, proposal, vesta, seam,
+Nearest to all: newton, reach, newly, campbell, absence, austin, history, coordinate,
+Nearest to used: phonetic, contents, mya, breeds, left, decline, disagree, appeal,
+Nearest to called: position, landing, attorney, railway, pi, entrance, split, vichy,
+Nearest to that: never, victoriae, and, habit, johan, photographs, reality, austin,
+Nearest to three: gb, mathbf, gland, tubing, nine, eight, zero, victoriae,
+Nearest to the: a, UNK, and, his, victoriae, mathbf, one, its,
+Nearest to when: interference, austin, suppress, eine, carnivores, mathbf, teachings, unfiltered,
+Nearest to its: the, accessed, version, striking, victoriae, pure, a, pierce,
+Nearest to use: gland, recommend, white, rigorously, accept, austin, politically, gettysburg,
+Nearest to he: it, mathbf, and, gollancz, gland, you, an, they,
+Nearest to six: nine, zero, mathbf, vs, phi, reginae, one, victoriae,
+Nearest to s: and, sahara, troops, currently, the, zero, boroughs, of,
+Average loss at step 12000 : 13.8088547783
+Average loss at step 14000 : 11.5112998215
+Average loss at step 16000 : 9.71534106708
+Average loss at step 18000 : 8.59706710231
+Average loss at step 20000 : 7.75074116433
+Nearest to a: the, this, agouti, or, adiabatic, and, culture, agnostic,
+Nearest to years: amsterdam, sciences, benzene, four, victoriae, six, five, happens,
+Nearest to may: zee, developers, psychology, polyhedra, extremophiles, precisely, meeting, koenigsegg,
+Nearest to new: circles, holiness, and, gb, proposal, offices, seam, vesta,
+Nearest to all: polyhedra, newton, history, ceiling, reach, coordinate, the, absence,
+Nearest to used: contents, breeds, phonetic, decline, disagree, mya, dasyprocta, left,
+Nearest to called: position, pi, attorney, exchanging, landing, entrance, railway, split,
+Nearest to that: which, habit, and, johan, never, victoriae, in, dasyprocta,
+Nearest to three: eight, five, two, six, seven, zero, nine, dasyprocta,
+Nearest to the: a, his, its, agouti, their, one, victoriae, adiabatic,
+Nearest to when: austin, interference, suppress, eine, carnivores, mathbf, bosnia, and,
+Nearest to its: the, his, their, aoc, tumor, accessed, metis, agouti,
+Nearest to use: ashoka, gland, rigorously, recommend, polyhedra, accept, shedding, gettysburg,
+Nearest to he: it, they, and, who, mathbf, there, was, zero,
+Nearest to six: nine, eight, zero, four, seven, three, five, mathbf,
+Nearest to s: and, zero, the, his, of, dasyprocta, sahara, troops,
+Average loss at step 22000 : 7.29002120793
+Average loss at step 24000 : 6.93063378406
+Average loss at step 26000 : 6.62375506437
+Average loss at step 28000 : 6.15877870786
+Average loss at step 30000 : 6.14214335346
+Nearest to a: the, this, agouti, or, aba, abet, victoriae, their,
+Nearest to years: amsterdam, sciences, benzene, four, victoriae, five, subtlety, happens,
+Nearest to may: zee, can, trudeau, developers, extremophiles, nine, koenigsegg, jack,
+Nearest to new: circles, holiness, yangon, seam, offices, gb, proposal, profound,
+Nearest to all: polyhedra, newton, history, absence, ceiling, virgin, coordinate, unfold,
+Nearest to used: contents, breeds, decline, disagree, left, phonetic, appeal, dasyprocta,
+Nearest to called: exchanging, pi, position, iic, attorney, entrance, champlain, split,
+Nearest to that: which, habit, never, johan, victoriae, polyhedra, antitrust, objective,
+Nearest to three: eight, four, five, six, seven, two, nine, zero,
+Nearest to the: their, its, his, a, agouti, adiabatic, this, victoriae,
+Nearest to when: primigenius, austin, suppress, and, eine, with, interference, mathbf,
+Nearest to its: the, their, his, aoc, metis, tumor, accessed, a,
+Nearest to use: ashoka, gland, rigorously, polyhedra, recommend, accept, gettysburg, rooms,
+Nearest to he: it, they, she, who, and, there, mathbf, gotten,
+Nearest to six: eight, nine, four, seven, five, three, zero, two,
+Nearest to s: and, his, of, the, zero, or, two, dasyprocta,
+Average loss at step 32000 : 5.85488548732
+Average loss at step 34000 : 5.83861638045
+Average loss at step 36000 : 5.72251720762
+Average loss at step 38000 : 5.25196716714
+Average loss at step 40000 : 5.48282074535
+Nearest to a: the, agouti, or, albury, this, aba, victoriae, abatis,
+Nearest to years: amsterdam, six, sciences, benzene, four, five, victoriae, reuptake,
+Nearest to may: can, zee, trudeau, should, will, eight, might, would,
+Nearest to new: circles, goo, albury, gb, monophosphate, holiness, profound, yangon,
+Nearest to all: polyhedra, newton, history, absence, coordinate, virgin, roper, mathbf,
+Nearest to used: decline, dasyprocta, breeds, contents, disagree, mya, monatomic, continuum,
+Nearest to called: exchanging, pi, position, iic, acetylcholine, entrance, champlain, split,
+Nearest to that: which, this, what, it, polyhedra, never, victoriae, habit,
+Nearest to three: four, six, five, eight, seven, two, zero, nine,
+Nearest to the: its, a, adiabatic, their, his, this, agouti, abet,
+Nearest to when: primigenius, austin, with, suppress, eine, including, and, but,
+Nearest to its: their, the, his, tumor, metis, aoc, agouti, abet,
+Nearest to use: ashoka, gland, polyhedra, recommend, rigorously, accept, gettysburg, rooms,
+Nearest to he: it, she, they, who, there, mathbf, but, gotten,
+Nearest to six: seven, eight, four, five, nine, three, zero, two,
+Nearest to s: and, his, dasyprocta, cinque, the, metis, two, kifl,
+Average loss at step 42000 : 5.32659191382
+Average loss at step 44000 : 5.27528748202
+Average loss at step 46000 : 5.24571657467
+Average loss at step 48000 : 5.05064962864
+Average loss at step 50000 : 5.15474711931
+Nearest to a: the, expired, agouti, hg, thibetanus, aba, appomattox, dipyramid,
+Nearest to years: amsterdam, four, six, hoax, sciences, happens, five, victoriae,
+Nearest to may: can, will, might, should, would, must, trudeau, eight,
+Nearest to new: circles, albury, naaman, seam, scientifically, holiness, goo, monophosphate,
+Nearest to all: polyhedra, newton, two, absence, mathbf, history, acts, three,
+Nearest to used: decline, dasyprocta, breeds, handicap, monatomic, found, known, disagree,
+Nearest to called: exchanging, iic, pi, position, acetylcholine, split, champlain, entrance,
+Nearest to that: which, what, this, never, polyhedra, victoriae, naaman, but,
+Nearest to three: four, six, seven, five, two, eight, one, nine,
+Nearest to the: its, their, his, agouti, adiabatic, a, this, victoriae,
+Nearest to when: primigenius, but, austin, eight, suppress, and, five, seven,
+Nearest to its: their, the, his, agouti, tumor, metis, aoc, accessed,
+Nearest to use: ashoka, polyhedra, gland, rigorously, recommend, accept, gettysburg, rooms,
+Nearest to he: it, she, they, who, there, this, gotten, mathbf,
+Nearest to six: eight, four, seven, five, three, nine, one, zero,
+Nearest to s: his, zero, dasyprocta, and, cinque, the, was, metis,
+Average loss at step 52000 : 5.18477400446
+Average loss at step 54000 : 5.11114428246
+Average loss at step 56000 : 5.03586944163
+Average loss at step 58000 : 5.17337947047
+Average loss at step 60000 : 4.93142760962
+Nearest to a: the, ssbn, wct, thibetanus, agouti, callithrix, cebus, aba,
+Nearest to years: four, amsterdam, six, microcebus, five, months, hoax, victoriae,
+Nearest to may: can, will, would, might, should, must, could, trudeau,
+Nearest to new: circles, naaman, albury, goo, seam, scientifically, callithrix, monophosphate,
+Nearest to all: microcebus, polyhedra, two, callithrix, cebus, acts, three, mathbf,
+Nearest to used: decline, handicap, microcebus, dasyprocta, found, continuum, known, cebus,
+Nearest to called: exchanging, pi, ssbn, acetylcholine, iic, split, champlain, tom,
+Nearest to that: which, this, what, never, cebus, it, objective, tamarin,
+Nearest to three: five, four, six, seven, two, eight, nine, one,
+Nearest to the: its, their, a, this, adiabatic, callithrix, agouti, his,
+Nearest to when: primigenius, after, austin, but, suppress, five, tamarin, and,
+Nearest to its: their, his, the, callithrix, her, agouti, cebus, metis,
+Nearest to use: ashoka, polyhedra, gland, rigorously, gettysburg, recommend, microcebus, expectation,
+Nearest to he: it, she, they, who, there, callithrix, but, ssbn,
+Nearest to six: eight, five, four, seven, nine, three, zero, callithrix,
+Nearest to s: his, wct, and, dasyprocta, callithrix, zero, was, cinque,
+Average loss at step 62000 : 4.80653275287
+Average loss at step 64000 : 4.80298928308
+Average loss at step 66000 : 4.9778061583
+Average loss at step 68000 : 4.92880989146
+Average loss at step 70000 : 4.78815213633
+Nearest to a: the, ssbn, upanija, wct, cebus, agouti, callithrix, expired,
+Nearest to years: four, five, months, amsterdam, microcebus, six, hoax, reuptake,
+Nearest to may: can, will, would, might, should, must, could, to,
+Nearest to new: circles, naaman, goo, albury, seam, scientifically, monophosphate, yangon,
+Nearest to all: many, tico, microcebus, some, polyhedra, callithrix, various, acts,
+Nearest to used: known, found, handicap, decline, microcebus, dasyprocta, cebus, agouti,
+Nearest to called: exchanging, ssbn, pi, champlain, acetylcholine, tom, ecc, iic,
+Nearest to that: which, what, this, never, cebus, objective, dinar, tamarin,
+Nearest to three: four, five, six, seven, two, eight, callithrix, nine,
+Nearest to the: its, their, this, callithrix, agouti, a, adiabatic, wct,
+Nearest to when: after, primigenius, austin, but, as, suppress, tamarin, with,
+Nearest to its: their, his, the, callithrix, her, cebus, metis, agouti,
+Nearest to use: polyhedra, ashoka, gland, unassigned, microcebus, callithrix, expectation, rigorously,
+Nearest to he: it, she, they, who, there, callithrix, but, ssbn,
+Nearest to six: eight, four, five, seven, three, nine, zero, callithrix,
+Nearest to s: wct, his, dasyprocta, zero, thz, cinque, callithrix, or,
+Average loss at step 72000 : 4.80325884366
+Average loss at step 74000 : 4.78602541548
+Average loss at step 76000 : 4.89614222682
+Average loss at step 78000 : 4.78169331312
+Average loss at step 80000 : 4.80316182685
+Nearest to a: the, ssbn, wct, upanija, cegep, thighs, dipyramid, agouti,
+Nearest to years: months, four, microcebus, amsterdam, hoax, five, reuptake, happens,
+Nearest to may: can, will, would, might, should, must, could, to,
+Nearest to new: circles, naaman, goo, seam, albury, scientifically, prague, monophosphate,
+Nearest to all: microcebus, many, two, tico, some, callithrix, polyhedra, these,
+Nearest to used: known, found, handicap, decline, dasyprocta, microcebus, agouti, cebus,
+Nearest to called: exchanging, ssbn, hood, champlain, protested, customization, pi, ecc,
+Nearest to that: which, what, this, objective, however, cebus, naaman, tamarin,
+Nearest to three: four, five, six, two, seven, eight, callithrix, one,
+Nearest to the: its, callithrix, their, agouti, wct, his, microsite, this,
+Nearest to when: after, clodius, austin, but, primigenius, tamarin, because, five,
+Nearest to its: their, his, the, her, callithrix, metis, agouti, tumor,
+Nearest to use: polyhedra, ashoka, gland, cegep, unassigned, accept, microcebus, callithrix,
+Nearest to he: it, she, they, who, there, callithrix, professions, iit,
+Nearest to six: five, four, eight, seven, three, nine, two, zero,
+Nearest to s: wct, dasyprocta, zero, his, cinque, callithrix, masterpieces, thz,
+Average loss at step 82000 : 4.80276642525
+Average loss at step 84000 : 4.78886307824
+Average loss at step 86000 : 4.73723094189
+Average loss at step 88000 : 4.69320081282
+Average loss at step 90000 : 4.75272560072
+Nearest to a: the, ssbn, wct, upanija, any, another, cegep, dipyramid,
+Nearest to years: months, microcebus, four, amsterdam, hoax, five, reuptake, vannevar,
+Nearest to may: can, will, would, might, should, must, could, cannot,
+Nearest to new: circles, naaman, goo, one, albury, thaler, gb, seam,
+Nearest to all: some, many, microcebus, tico, these, callithrix, various, polyhedra,
+Nearest to used: known, found, microcebus, handicap, dasyprocta, seen, agouti, decline,
+Nearest to called: exchanging, hood, ssbn, split, customization, thaler, protested, champlain,
+Nearest to that: which, what, however, this, but, cebus, dinar, tamarin,
+Nearest to three: four, two, five, seven, eight, six, callithrix, cegep,
+Nearest to the: its, their, callithrix, agouti, wct, adiabatic, a, cegep,
+Nearest to when: after, clodius, but, if, before, tamarin, because, primigenius,
+Nearest to its: their, his, the, her, callithrix, agouti, metis, wct,
+Nearest to use: polyhedra, ashoka, gland, cegep, callithrix, microcebus, clodius, catalysis,
+Nearest to he: she, it, they, there, who, callithrix, iit, zero,
+Nearest to six: seven, eight, five, four, nine, three, zero, callithrix,
+Nearest to s: wct, dasyprocta, his, thz, chalcedon, cinque, zero, mating,
+Average loss at step 92000 : 4.71442663682
+Average loss at step 94000 : 4.60804726839
+Average loss at step 96000 : 4.72456447947
+Average loss at step 98000 : 4.62782734013
+Average loss at step 100000 : 4.67684453142
+Nearest to a: the, ssbn, any, upanija, wct, cegep, another, expired,
+Nearest to years: months, four, microcebus, amsterdam, days, hoax, reuptake, happens,
+Nearest to may: can, will, would, might, should, could, must, cannot,
+Nearest to new: circles, naaman, goo, seam, prague, albury, scientifically, thaler,
+Nearest to all: many, some, microcebus, tico, these, various, callithrix, two,
+Nearest to used: known, found, dasyprocta, cebus, microcebus, handicap, seen, agouti,
+Nearest to called: exchanging, hood, customization, thaler, split, protested, ssbn, ecc,
+Nearest to that: which, what, however, this, but, polyhedra, cebus, cegep,
+Nearest to three: four, five, six, two, seven, eight, callithrix, nine,
+Nearest to the: its, their, callithrix, nordisk, adiabatic, agouti, wct, his,
+Nearest to when: after, if, clodius, before, but, because, tamarin, where,
+Nearest to its: their, his, the, her, callithrix, agouti, wct, thz,
+Nearest to use: polyhedra, ashoka, cegep, callithrix, microcebus, gland, catalysis, clodius,
+Nearest to he: she, it, they, who, there, callithrix, iit, already,
+Nearest to six: seven, eight, five, four, nine, three, two, callithrix,
+Nearest to s: wct, his, dasyprocta, thz, the, was, chalcedon, callithrix,
+/usr/lib64/python2.7/site-packages/matplotlib/collections.py:590: FutureWarning: elementwise comparison failed; returning scalar instead, but in the future will perform elementwise comparison
+ if self._edgecolors == str('face'):
+
diff --git a/hw11_natural_language_processing/martink_word2vec/text8.zip b/hw11_natural_language_processing/martink_word2vec/text8.zip
new file mode 100644
index 0000000000000000000000000000000000000000..436e05b2dffe4f084fdada65374bd7804bdce097
Binary files /dev/null and b/hw11_natural_language_processing/martink_word2vec/text8.zip differ
diff --git a/hw11_natural_language_processing/martink_word2vec/tsne.png b/hw11_natural_language_processing/martink_word2vec/tsne.png
new file mode 100644
index 0000000000000000000000000000000000000000..a62f3035d1470d3dfa7cf4b116408e1e340bdbcc
Binary files /dev/null and b/hw11_natural_language_processing/martink_word2vec/tsne.png differ
diff --git a/hw11_natural_language_processing/martink_word2vec/word2vec_basic.py b/hw11_natural_language_processing/martink_word2vec/word2vec_basic.py
new file mode 100644
index 0000000000000000000000000000000000000000..c717693a567249ea00c96379e58ba4aeb5ed9f8d
--- /dev/null
+++ b/hw11_natural_language_processing/martink_word2vec/word2vec_basic.py
@@ -0,0 +1,249 @@
+# Copyright 2015 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import collections
+import math
+import os
+import random
+import zipfile
+
+import numpy as np
+from six.moves import urllib
+from six.moves import xrange # pylint: disable=redefined-builtin
+import tensorflow as tf
+
+# Step 1: Download the data.
+url = 'http://mattmahoney.net/dc/'
+
+
+def maybe_download(filename, expected_bytes):
+ """Download a file if not present, and make sure it's the right size."""
+ if not os.path.exists(filename):
+ filename, _ = urllib.request.urlretrieve(url + filename, filename)
+ statinfo = os.stat(filename)
+ if statinfo.st_size == expected_bytes:
+ print('Found and verified', filename)
+ else:
+ print(statinfo.st_size)
+ raise Exception(
+ 'Failed to verify ' + filename + '. Can you get to it with a browser?')
+ return filename
+
+filename = maybe_download('text8.zip', 31344016)
+
+
+# Read the data into a list of strings.
+def read_data(filename):
+ """Extract the first file enclosed in a zip file as a list of words"""
+ with zipfile.ZipFile(filename) as f:
+ data = tf.compat.as_str(f.read(f.namelist()[0])).split()
+ return data
+
+words = read_data(filename)
+print('Data size', len(words))
+
+# Step 2: Build the dictionary and replace rare words with UNK token.
+vocabulary_size = 50000
+
+
+def build_dataset(words):
+ count = [['UNK', -1]]
+ count.extend(collections.Counter(words).most_common(vocabulary_size - 1))
+ dictionary = dict()
+ for word, _ in count:
+ dictionary[word] = len(dictionary)
+ data = list()
+ unk_count = 0
+ for word in words:
+ if word in dictionary:
+ index = dictionary[word]
+ else:
+ index = 0 # dictionary['UNK']
+ unk_count += 1
+ data.append(index)
+ count[0][1] = unk_count
+ reverse_dictionary = dict(zip(dictionary.values(), dictionary.keys()))
+ return data, count, dictionary, reverse_dictionary
+
+data, count, dictionary, reverse_dictionary = build_dataset(words)
+del words # Hint to reduce memory.
+print('Most common words (+UNK)', count[:5])
+print('Sample data', data[:10], [reverse_dictionary[i] for i in data[:10]])
+
+data_index = 0
+
+
+# Step 3: Function to generate a training batch for the skip-gram model.
+def generate_batch(batch_size, num_skips, skip_window):
+ global data_index
+ assert batch_size % num_skips == 0
+ assert num_skips <= 2 * skip_window
+ batch = np.ndarray(shape=(batch_size), dtype=np.int32)
+ labels = np.ndarray(shape=(batch_size, 1), dtype=np.int32)
+ span = 2 * skip_window + 1 # [ skip_window target skip_window ]
+ buffer = collections.deque(maxlen=span)
+ for _ in range(span):
+ buffer.append(data[data_index])
+ data_index = (data_index + 1) % len(data)
+ for i in range(batch_size // num_skips):
+ target = skip_window # target label at the center of the buffer
+ targets_to_avoid = [skip_window]
+ for j in range(num_skips):
+ while target in targets_to_avoid:
+ target = random.randint(0, span - 1)
+ targets_to_avoid.append(target)
+ batch[i * num_skips + j] = buffer[skip_window]
+ labels[i * num_skips + j, 0] = buffer[target]
+ buffer.append(data[data_index])
+ data_index = (data_index + 1) % len(data)
+ return batch, labels
+
+batch, labels = generate_batch(batch_size=8, num_skips=2, skip_window=1)
+for i in range(8):
+ print(batch[i], reverse_dictionary[batch[i]],
+ '->', labels[i, 0], reverse_dictionary[labels[i, 0]])
+
+# Step 4: Build and train a skip-gram model.
+
+batch_size = 128
+embedding_size = 128 # Dimension of the embedding vector.
+skip_window = 1 # How many words to consider left and right.
+num_skips = 2 # How many times to reuse an input to generate a label.
+
+# We pick a random validation set to sample nearest neighbors. Here we limit the
+# validation samples to the words that have a low numeric ID, which by
+# construction are also the most frequent.
+valid_size = 16 # Random set of words to evaluate similarity on.
+valid_window = 100 # Only pick dev samples in the head of the distribution.
+valid_examples = np.random.choice(valid_window, valid_size, replace=False)
+num_sampled = 64 # Number of negative examples to sample.
+
+graph = tf.Graph()
+
+with graph.as_default():
+
+ # Input data.
+ train_inputs = tf.placeholder(tf.int32, shape=[batch_size])
+ train_labels = tf.placeholder(tf.int32, shape=[batch_size, 1])
+ valid_dataset = tf.constant(valid_examples, dtype=tf.int32)
+
+ # Ops and variables pinned to the CPU because of missing GPU implementation
+ with tf.device('/cpu:0'):
+ # Look up embeddings for inputs.
+ embeddings = tf.Variable(
+ tf.random_uniform([vocabulary_size, embedding_size], -1.0, 1.0))
+ embed = tf.nn.embedding_lookup(embeddings, train_inputs)
+
+ # Construct the variables for the NCE loss
+ nce_weights = tf.Variable(
+ tf.truncated_normal([vocabulary_size, embedding_size],
+ stddev=1.0 / math.sqrt(embedding_size)))
+ nce_biases = tf.Variable(tf.zeros([vocabulary_size]))
+
+ # Compute the average NCE loss for the batch.
+ # tf.nce_loss automatically draws a new sample of the negative labels each
+ # time we evaluate the loss.
+ loss = tf.reduce_mean(
+ tf.nn.nce_loss(nce_weights, nce_biases, embed, train_labels,
+ num_sampled, vocabulary_size))
+
+ # Construct the SGD optimizer using a learning rate of 1.0.
+ optimizer = tf.train.GradientDescentOptimizer(1.0).minimize(loss)
+
+ # Compute the cosine similarity between minibatch examples and all embeddings.
+ norm = tf.sqrt(tf.reduce_sum(tf.square(embeddings), 1, keep_dims=True))
+ normalized_embeddings = embeddings / norm
+ valid_embeddings = tf.nn.embedding_lookup(
+ normalized_embeddings, valid_dataset)
+ similarity = tf.matmul(
+ valid_embeddings, normalized_embeddings, transpose_b=True)
+
+ # Add variable initializer.
+ init = tf.initialize_all_variables()
+
+# Step 5: Begin training.
+num_steps = 100001
+
+with tf.Session(graph=graph) as session:
+ # We must initialize all variables before we use them.
+ init.run()
+ print("Initialized")
+
+ average_loss = 0
+ for step in xrange(num_steps):
+ batch_inputs, batch_labels = generate_batch(
+ batch_size, num_skips, skip_window)
+ feed_dict = {train_inputs: batch_inputs, train_labels: batch_labels}
+
+ # We perform one update step by evaluating the optimizer op (including it
+ # in the list of returned values for session.run()
+ _, loss_val = session.run([optimizer, loss], feed_dict=feed_dict)
+ average_loss += loss_val
+
+ if step % 2000 == 0:
+ if step > 0:
+ average_loss /= 2000
+ # The average loss is an estimate of the loss over the last 2000 batches.
+ print("Average loss at step ", step, ": ", average_loss)
+ average_loss = 0
+
+ # Note that this is expensive (~20% slowdown if computed every 500 steps)
+ if step % 10000 == 0:
+ sim = similarity.eval()
+ for i in xrange(valid_size):
+ valid_word = reverse_dictionary[valid_examples[i]]
+ top_k = 8 # number of nearest neighbors
+ nearest = (-sim[i, :]).argsort()[1:top_k + 1]
+ log_str = "Nearest to %s:" % valid_word
+ for k in xrange(top_k):
+ close_word = reverse_dictionary[nearest[k]]
+ log_str = "%s %s," % (log_str, close_word)
+ print(log_str)
+ final_embeddings = normalized_embeddings.eval()
+
+# Step 6: Visualize the embeddings.
+
+
+def plot_with_labels(low_dim_embs, labels, filename='tsne.png'):
+ assert low_dim_embs.shape[0] >= len(labels), "More labels than embeddings"
+ plt.figure(figsize=(18, 18)) # in inches
+ for i, label in enumerate(labels):
+ x, y = low_dim_embs[i, :]
+ plt.scatter(x, y)
+ plt.annotate(label,
+ xy=(x, y),
+ xytext=(5, 2),
+ textcoords='offset points',
+ ha='right',
+ va='bottom')
+
+ plt.savefig(filename)
+
+try:
+ from sklearn.manifold import TSNE
+ import matplotlib.pyplot as plt
+
+ tsne = TSNE(perplexity=30, n_components=2, init='pca', n_iter=5000)
+ plot_only = 500
+ low_dim_embs = tsne.fit_transform(final_embeddings[:plot_only, :])
+ labels = [reverse_dictionary[i] for i in xrange(plot_only)]
+ plot_with_labels(low_dim_embs, labels)
+
+except ImportError:
+ print("Please install sklearn, matplotlib, and scipy to visualize embeddings.")