Code

[MRG] Poincare Model implementation#1696

Code

Merged

menshikh-iv merged 116 commits intopoincarepiskvorky/gensim:poincarefrom poincare_modelpiskvorky/gensim:poincare_modelCopy head branch name to clipboardNov 15, 2017

+1,181-1

Commits

Commits on Oct 26, 2017

• Initial classes and loading data for poincare model

  jayantj

  committed
  6afdd22

  Copy full SHA for 6afdd22

• Initial implementation of training using autograd

  jayantj

  committed
  a804006

  Copy full SHA for a804006

• faster negative sampling, bugfix in vector updates

  jayantj

  committed
  6bd0d4b

  Copy full SHA for 6bd0d4b

• allows poincare dist function to be differentiable by autograd

  jayantj

  committed
  98f94a7

  Copy full SHA for 98f94a7

• batched gradient descent initial implementation

  jayantj

  committed
  b727523

  Copy full SHA for b727523

• minor changes to batch poincare distance computation

  jayantj

  committed
  1e6aee1

  Copy full SHA for 1e6aee1

• Adds calculation of gradients for poincare model

  jayantj

  committed
  e286a0b

  Copy full SHA for e286a0b

• Correct implementation of clipping of updated vectors

  jayantj

  committed
  3e28e8b

  Copy full SHA for 3e28e8b

• Fixes error in gradient computation

  jayantj

  committed
  99a2270

  Copy full SHA for 99a2270

• Better messages while training

  jayantj

  committed
  2e9e31c

  Copy full SHA for 2e9e31c

Commits on Oct 27, 2017

• Renames PoincareDistance to PoincareExample for clarity

  jayantj

  committed
  d72cb10

  Copy full SHA for d72cb10

• Compares computed gradients to autograd gradients every few iterations

  jayantj

  committed
  d439501

  Copy full SHA for d439501

• Avoids doing some numpy computations twice

  jayantj

  committed
  e1ed24d

  Copy full SHA for e1ed24d

• Avoids creating copies of numpy vectors

  jayantj

  committed
  3b2a383

  Copy full SHA for 3b2a383

• Only calls nan_to_num when gamma has at least one value equal to 1

  jayantj

  committed
  7d68aae

  Copy full SHA for 7d68aae

• Simply sets nan gradients to zero instead of nan_to_num

  jayantj

  committed
  ba82d42

  Copy full SHA for ba82d42

• Adds batch-wise implementation of training and gradient computations

  jayantj

  committed
  71f61d1

  Copy full SHA for 71f61d1

Commits on Oct 30, 2017

• Minor correction in clipping

  jayantj

  committed
  2a5a7fb

  Copy full SHA for 2a5a7fb

• Merge branch 'poincare' into poincare_model

  jayantj

  committed
  0c57aa1

  Copy full SHA for 0c57aa1

• Fixes typo in clip_vectors

  jayantj

  committed
  9c51609

  Copy full SHA for 9c51609

Commits on Oct 31, 2017

• Prints average loss every few iterations instead of current loss

  jayantj

  committed
  f22d9b2

  Copy full SHA for f22d9b2

Commits on Nov 2, 2017

• Adds weighted negative sampling

  jayantj

  committed
  7905c8c

  Copy full SHA for 7905c8c

• Ensures positive edges are not returned by negative sampling

  jayantj

  committed
  075df25

  Copy full SHA for 075df25

• Poincare model stores node indices in relations instead of node keys

  jayantj

  committed
  6060e56

  Copy full SHA for 6060e56

• Minor renaming; uses node indices for batch training instead of node keys

  jayantj

  committed
  8ea8f23

  Copy full SHA for 8ea8f23

Commits on Nov 3, 2017

• Changes shapes of vectors passed to PoincareBatch

  jayantj

  committed
  b8d77e3

  Copy full SHA for b8d77e3

• Minor bugfixes related to batch size

  jayantj

  committed
  0011b93

  Copy full SHA for 0011b93

• Corrects implementation of negative sampling for batch training

  jayantj

  committed
  b52ee2e

  Copy full SHA for b52ee2e

• Adds option to check gradients in batchwise training

  jayantj

  committed
  d247384

  Copy full SHA for d247384

• Checks gradients only every few iterations

  jayantj

  committed
  8c4f5a3

  Copy full SHA for 8c4f5a3

• Handles multiple occurrence of same node across and within batches

  jayantj

  committed
  34b0ad3

  Copy full SHA for 34b0ad3

• Removes unused section of code

  jayantj

  committed
  1779cd7

  Copy full SHA for 1779cd7

• Implements slightly different clipping method

  jayantj

  committed
  faacb43

  Copy full SHA for faacb43

• Fixes bugs with wrong reshape in batchwise training

  jayantj

  committed
  c68088e

  Copy full SHA for c68088e

• Example-wise training takes into account multiple occurrences of same node in an example too

  jayantj

  committed
  0c2f2cb

  Copy full SHA for 0c2f2cb

• Batchwise training prints average loss over many iterations instead of current batch

  jayantj

  committed
  386f602

  Copy full SHA for 386f602

• Fixes bug in updating vector for batchwise training

  jayantj

  committed
  f0fb9e9

  Copy full SHA for f0fb9e9

• Faster implementation of negative sampling

  jayantj

  committed
  7d8fbec

  Copy full SHA for 7d8fbec

Commits on Nov 6, 2017

• Negative sampling for a node follows different paths depending on fraction of positive relations

  jayantj

  committed
  315f95c

  Copy full SHA for 315f95c

• Uses a buffer for negative samples to reduce calls to np.random.choice

  jayantj

  committed
  0802dd5

  Copy full SHA for 0802dd5

• Cleans up poincare.py, removes unused code

  jayantj

  committed
  a106191

  Copy full SHA for a106191

• Adds shapes to PoincareBatch, more documentation

  jayantj

  committed
  1aa586d

  Copy full SHA for 1aa586d

• Adds more documentation to PoincareModel

  jayantj

  committed
  13b00dc

  Copy full SHA for 13b00dc

• Stores indices for nodes in a batch in PoincareBatch for better encapsulation

  jayantj

  committed
  5978af6

  Copy full SHA for 5978af6

• More documentation for poincare module

  jayantj

  committed
  e40c3e3

  Copy full SHA for e40c3e3

• Implements burn-in for poincare model

  jayantj

  committed
  ec8b516

  Copy full SHA for ec8b516

• Slightly better logging for poincare model

  jayantj

  committed
  86ae4d6

  Copy full SHA for 86ae4d6

Commits on Nov 7, 2017

• Uses np.random.random and np.searchsorted for random sampling rather than np.random.choice

  jayantj

  committed
  ac51e9c

  Copy full SHA for ac51e9c

• Removes duplicates in negative samples

  jayantj

  committed
  5900c6f

  Copy full SHA for 5900c6f

• Moves helper classes in poincare after PoincareModel

  jayantj

  committed
  4ac4d2e

  Copy full SHA for 4ac4d2e

• Change in PoincareModel API to allow initializing from an iterable, separate class for streaming from file

  jayantj

  committed
  9eb6f48

  Copy full SHA for 9eb6f48

• Adds failing test for handling encoding in PoincareData

  jayantj

  committed
  2ded72b

  Copy full SHA for 2ded72b

• Fixes encoding handling in PoincareData

  jayantj

  committed
  81960e1

  Copy full SHA for 81960e1

• Adds docstrings to PoincareData, PoincareData streams tuples now

  jayantj

  committed
  5de194b

  Copy full SHA for 5de194b

Commits on Nov 8, 2017

• More unittests for PoincareModel

  jayantj

  committed
  6dd6915

  Copy full SHA for 6dd6915

• Changes handle_duplicates to staticmethod, adds test

  jayantj

  committed
  4b502af

  Copy full SHA for 4b502af

• Adds batch size and print_every parameters to train method

  jayantj

  committed
  12be121

  Copy full SHA for 12be121

• Renames print_check to should_print

  jayantj

  committed
  29e799c

  Copy full SHA for 29e799c

• Adds separate parameter for checking gradients

  jayantj

  committed
  b4ff1dd

  Copy full SHA for b4ff1dd

• Minor fixes for coding style

  jayantj

  committed
  e2f72bc

  Copy full SHA for e2f72bc

• Removes default values from docstrings, redundant

  jayantj

  committed
  953b4a7

  Copy full SHA for 953b4a7

• Adds example to PoincareModel init docstring

  jayantj

  committed
  eebc12a

  Copy full SHA for eebc12a

• Extracts buffer for negatives out into a separate class

  jayantj

  committed
  21a1c82

  Copy full SHA for 21a1c82

• More detailed logging, fix to check_gradients

  jayantj

  committed
  f9325ea

  Copy full SHA for f9325ea

• Minor fixes to documentation in poincare.py

  jayantj

  committed
  5db8456

  Copy full SHA for 5db8456

• Adds tests for gradients checking

  jayantj

  committed
  53030a0

  Copy full SHA for 53030a0

• Raise AssertionError if gradients check fails

  jayantj

  committed
  db0d293

  Copy full SHA for db0d293

• Adds failing tests for saving/loading PoincareModel instances

  jayantj

  committed
  1adf81a

  Copy full SHA for 1adf81a

• Fixes bug with saving/loading PoincareModel to disk

  jayantj

  committed
  3898089

  Copy full SHA for 3898089

• Adds test and fix for raising error on invalid input data

  jayantj

  committed
  5cd913a

  Copy full SHA for 5cd913a

• Adds test and fix for no duplicates and positives in negative sample

  jayantj

  committed
  6305228

  Copy full SHA for 6305228

• Bugfix with NegativesBuffer having less than items left

  jayantj

  committed
  fb13eb5

  Copy full SHA for fb13eb5

• Uses larger data for poincare tests, adds data files

  jayantj

  committed
  ea2fd48

  Copy full SHA for ea2fd48

• Bugfix with incorrect use of random state

  jayantj

  committed
  110fb1e

  Copy full SHA for 110fb1e

Commits on Nov 9, 2017

• Minor fixes in documentation style

  jayantj

  committed
  0aeec2f

  Copy full SHA for 0aeec2f

• Renames PoincareData to PoincareRelations

  jayantj

  committed
  38feb7a

  Copy full SHA for 38feb7a

• Change in the order of conditions checked before resampling

  jayantj

  committed
  0e7ebb3

  Copy full SHA for 0e7ebb3

• Merge branch 'poincare' into poincare_model

  jayantj

  committed
  52a1e57

  Copy full SHA for 52a1e57

• Imports datapath from test.utils instead of defining own

  jayantj

  committed
  630771d

  Copy full SHA for 630771d

• Adds working examples and a more detailed description in docstring

  jayantj

  committed
  7c6d972

  Copy full SHA for 7c6d972

• Renames term_relations to node_relations

  jayantj

  committed
  16dcf0b

  Copy full SHA for 16dcf0b

• Removes unused imports

  jayantj

  committed
  3501d6f

  Copy full SHA for 3501d6f

• Moves iter parameter to train instead of __init__, renames to epochs

  jayantj

  committed
  d690a25

  Copy full SHA for d690a25

• Fixes term_relations in tests

  jayantj

  committed
  2383e82

  Copy full SHA for 2383e82

• Adds option to disable gradient check, disabled by default

  jayantj

  committed
  3ed0bea

  Copy full SHA for 3ed0bea

• Extracts gradient checking code into a separate method

  jayantj

  committed
  9f562cb

  Copy full SHA for 9f562cb

• Conditionally import autograd only if gradient checking is enabled

  jayantj

  committed
  98e078d

  Copy full SHA for 98e078d

• Marks private methods in poincare module with leading underscore

  jayantj

  committed
  530146d

  Copy full SHA for 530146d

• Adds init_range as an API parameter to PoincareModel

  jayantj

  committed
  d17c075

  Copy full SHA for d17c075

• Marks private properties with a leading underscore

  jayantj

  committed
  be0249a

  Copy full SHA for be0249a

• Fixes bug with burn-in happening on subsequent calls to train

  jayantj

  committed
  dc2ab95

  Copy full SHA for dc2ab95

• Adds test for training multiple times

  jayantj

  committed
  a306f20

  Copy full SHA for a306f20

• Adds autograd to test dependencies

  jayantj

  committed
  f9750e6

  Copy full SHA for f9750e6

• Renames wv to kv in PoincareModel

  jayantj

  committed
  b7212ff

  Copy full SHA for b7212ff

Commits on Nov 10, 2017

• add numpy==1.12 as test dependency

  menshikh-iv

  authored
  3556ee4

  Copy full SHA for 3556ee4

• add missing quote

  menshikh-iv

  authored
  4644eda

  Copy full SHA for 4644eda

Commits on Nov 13, 2017

• try to run tests without autograd

  menshikh-iv

  authored
  94a2a18

  Copy full SHA for 94a2a18

• fix PEP8 in poincare.py

  menshikh-iv

  authored
  7a4ec79

  Copy full SHA for 7a4ec79

• fix PEP8 in test_poincare

  menshikh-iv

  authored
  613ca38

  Copy full SHA for 613ca38

• PoincareRelations handles python2 correctly

  jayantj

  committed
  3029d41

  Copy full SHA for 3029d41

• Bugfix with int division for python2

  jayantj

  committed
  055044c

  Copy full SHA for 055044c

• Imports mock module for tests correctly in python2

  jayantj

  committed
  f75491f

  Copy full SHA for f75491f

• Cleaner implementation of __iter__ for PoincareRelations

  jayantj

  committed
  59fcf8b

  Copy full SHA for 59fcf8b

• Adds rst file and updates apiref.rst for poincare module

  jayantj

  committed
  dcbe7aa

  Copy full SHA for dcbe7aa

• Adds clarifying comment to PoincareRelations.__iter__

  jayantj

  committed
  b69f51f

  Copy full SHA for b69f51f

• Updates rst file for poincare

  jayantj

  committed
  001ec76

  Copy full SHA for 001ec76

• Renames hypernym pair to relations everywhere

  jayantj

  committed
  9446a05

  Copy full SHA for 9446a05

• Simpler way of detecting duplicates

  jayantj

  committed
  930dfd4

  Copy full SHA for 930dfd4

Commits on Nov 14, 2017

• Minor documentation updates in poincare.py

  jayantj

  committed
  355e521

  Copy full SHA for 355e521

• Skips gradients test if autograd not installed, adds test for bytes input data

  jayantj

  committed
  0d5175c

  Copy full SHA for 0d5175c

• Fix flake8 (noqa + remove unused var)

  menshikh-iv

  committed
  00ca7ab

  Copy full SHA for 00ca7ab

• Fix missing mock dependency for win

  menshikh-iv

  committed
  8ff23ae

  Copy full SHA for 8ff23ae

• Fix links in docstrings

  menshikh-iv

  committed
  30ac3e6

  Copy full SHA for 30ac3e6

Commits on Nov 15, 2017

• Changes error message for negative sampling failing

  jayantj

  committed
  a928ca1

  Copy full SHA for a928ca1

• Adds option to specify dtype for PoincareModel and corresponding unittest

  jayantj

  committed
  dfc19cb

  Copy full SHA for dfc19cb

• Extends test for dtype to check after training, updates docstring

  jayantj

  committed
  e967c54

  Copy full SHA for e967c54