Commit Graph

861 Commits

Author SHA1 Message Date
Matthew Honnibal
95a9615221 Fix loading of multiple pre-trained vectors
This patch addresses #1660, which was caused by keying all pre-trained
vectors with the same ID when telling Thinc how to refer to them. This
meant that if multiple models were loaded that had pre-trained vectors,
errors or incorrect behaviour resulted.

The vectors class now includes a .name attribute, which defaults to:
{nlp.meta['lang']_nlp.meta['name']}.vectors
The vectors name is set in the cfg of the pipeline components under the
key pretrained_vectors. This replaces the previous cfg key
pretrained_dims.

In order to make existing models compatible with this change, we check
for the pretrained_dims key when loading models in from_disk and
from_bytes, and add the cfg key pretrained_vectors if we find it.
2018-03-28 16:02:59 +02:00
Matthew Honnibal
18da89e04c Handle non-callable gold_tuples in parser begin_training 2018-03-27 21:08:41 +02:00
Matthew Honnibal
1f7229f40f Revert "Merge branch 'develop' of https://github.com/explosion/spaCy into develop"
This reverts commit c9ba3d3c2d, reversing
changes made to 92c26a35d4.
2018-03-27 19:23:02 +02:00
Matthew Honnibal
f57bfbccdc Fix non-projective label filtering 2018-03-27 13:41:33 +02:00
Matthew Honnibal
d2118792e7 Merge changes from master 2018-03-27 13:38:41 +02:00
Matthew Honnibal
25280b7013 Try to make sum_state_features faster 2018-03-27 10:08:38 +00:00
Matthew Honnibal
987e1533a4 Use 8 features in parser 2018-03-27 10:08:12 +00:00
Matthew Honnibal
dd54511c4f Pass data as a function in begin_training methods 2018-03-27 09:39:59 +00:00
Matthew Honnibal
d9ebd78e11 Change default sizes in parser 2018-03-26 17:22:18 +02:00
Matthew Honnibal
49fbe2dfee Use thinc.openblas in spacy.syntax.nn_parser 2018-03-20 02:22:09 +01:00
Matthew Honnibal
bede11b67c
Improve label management in parser and NER (#2108)
This patch does a few smallish things that tighten up the training workflow a little, and allow memory use during training to be reduced by letting the GoldCorpus stream data properly.

Previously, the parser and entity recognizer read and saved labels as lists, with extra labels noted separately. Lists were used becaue ordering is very important, to ensure that the label-to-class mapping is stable.

We now manage labels as nested dictionaries, first keyed by the action, and then keyed by the label. Values are frequencies. The trick is, how do we save new labels? We need to make sure we iterate over these in the same order they're added. Otherwise, we'll get different class IDs, and the model's predictions won't make sense.

To allow stable sorting, we map the new labels to negative values. If we have two new labels, they'll be noted as having "frequency" -1 and -2. The next new label will then have "frequency" -3. When we sort by (frequency, label), we then get a stable sort.

Storing frequencies then allows us to make the next nice improvement. Previously we had to iterate over the whole training set, to pre-process it for the deprojectivisation. This led to storing the whole training set in memory. This was most of the required memory during training.

To prevent this, we now store the frequencies as we stream in the data, and deprojectivize as we go. Once we've built the frequencies, we can then apply a frequency cut-off when we decide how many classes to make.

Finally, to allow proper data streaming, we also have to have some way of shuffling the iterator. This is awkward if the training files have multiple documents in them. To solve this, the GoldCorpus class now writes the training data to disk in msgpack files, one per document. We can then shuffle the data by shuffling the paths.

This is a squash merge, as I made a lot of very small commits. Individual commit messages below.

* Simplify label management for TransitionSystem and its subclasses

* Fix serialization for new label handling format in parser

* Simplify and improve GoldCorpus class. Reduce memory use, write to temp dir

* Set actions in transition system

* Require thinc 6.11.1.dev4

* Fix error in parser init

* Add unicode declaration

* Fix unicode declaration

* Update textcat test

* Try to get model training on less memory

* Print json loc for now

* Try rapidjson to reduce memory use

* Remove rapidjson requirement

* Try rapidjson for reduced mem usage

* Handle None heads when projectivising

* Stream json docs

* Fix train script

* Handle projectivity in GoldParse

* Fix projectivity handling

* Add minibatch_by_words util from ud_train

* Minibatch by number of words in spacy.cli.train

* Move minibatch_by_words util to spacy.util

* Fix label handling

* More hacking at label management in parser

* Fix encoding in msgpack serialization in GoldParse

* Adjust batch sizes in parser training

* Fix minibatch_by_words

* Add merge_subtokens function to pipeline.pyx

* Register merge_subtokens factory

* Restore use of msgpack tmp directory

* Use minibatch-by-words in train

* Handle retokenization in scorer

* Change back-off approach for missing labels. Use 'dep' label

* Update NER for new label management

* Set NER tags for over-segmented words

* Fix label alignment in gold

* Fix label back-off for infrequent labels

* Fix int type in labels dict key

* Fix int type in labels dict key

* Update feature definition for 8 feature set

* Update ud-train script for new label stuff

* Fix json streamer

* Print the line number if conll eval fails

* Update children and sentence boundaries after deprojectivisation

* Export set_children_from_heads from doc.pxd

* Render parses during UD training

* Remove print statement

* Require thinc 6.11.1.dev6. Try adding wheel as install_requires

* Set different dev version, to flush pip cache

* Update thinc version

* Update GoldCorpus docs

* Remove print statements

* Fix formatting and links [ci skip]
2018-03-19 02:58:08 +01:00
Matthew Honnibal
307d6bf6d3 Fix parser for Thinc 6.11 2018-03-16 10:59:31 +01:00
Matthew Honnibal
9a389c4490 Fix parser for Thinc 6.11 2018-03-16 10:38:13 +01:00
Matthew Honnibal
648532d647 Don't assume blas methods are present 2018-03-16 02:48:20 +01:00
Matthew Honnibal
e101f10ef0 Fix header 2018-03-13 02:12:16 +01:00
Matthew Honnibal
d55620041b Switch parser to gemm from thinc.openblas 2018-03-13 02:10:58 +01:00
Matthew Honnibal
4b72c38556 Fix dropout bug in beam parser 2018-03-10 23:16:40 +01:00
Matthew Honnibal
3d6487c734 Support dropout in beam parse 2018-03-10 22:41:55 +01:00
Matthew Honnibal
14f729c72a Add subtok label to parser 2018-02-26 12:26:35 +01:00
Matthew Honnibal
7137ad8b0b Make label filtering clearer for projectivisation 2018-02-26 12:02:01 +01:00
Matthew Honnibal
7b66ec896a Revert "Revert "Improve parser oracle around sentence breaks.""
This reverts commit 36e481c584.
2018-02-26 10:57:37 +01:00
Matthew Honnibal
36e481c584 Revert "Improve parser oracle around sentence breaks."
This reverts commit 50817dc9ad.
2018-02-26 10:53:55 +01:00
Matthew Honnibal
50817dc9ad Improve parser oracle around sentence breaks. 2018-02-22 19:22:26 +01:00
Matthew Honnibal
661873ee4c Randomize the rebatch size in parser 2018-02-21 21:02:07 +01:00
Matthew Honnibal
a0ddb803fd Make error when no label found more helpful 2018-02-21 16:00:59 +01:00
Matthew Honnibal
ea2fc5d45f Improve length and freq cutoffs in parser 2018-02-21 16:00:38 +01:00
Matthew Honnibal
e5757d4bf0 Add labels property to parser 2018-02-21 16:00:00 +01:00
Matthew Honnibal
eff4ae809a Fix nonproj label filter 2018-02-21 15:59:04 +01:00
Matthew Honnibal
e624405cda Temporarily remove cutoff when filtering labels in nonproj 2018-02-21 13:53:40 +01:00
Matthew Honnibal
8f06903e09 Fix multitask objectives 2018-02-17 18:41:36 +01:00
Matthew Honnibal
d1246c95fb Fix model loading when using multitask objectives 2018-02-17 18:11:36 +01:00
Matthew Honnibal
7d5c720fc3 Fix multitask objective when no pipeline provided 2018-02-15 23:50:21 +01:00
Matthew Honnibal
59b7cf9db8 Add get_beam_parse method in ArcEager, for Prodigy 2018-02-15 21:03:16 +01:00
Claudiu-Vlad Ursache
e28de12cbd
Ensure files opened in from_disk are closed
Fixes [issue 1706](https://github.com/explosion/spaCy/issues/1706).
2018-02-13 20:49:43 +01:00
Matthew Honnibal
e361b4f82b Fix #1929: Incorrect NER when pre-set sentence boundaries. 2018-02-08 15:25:41 +01:00
Matthew Honnibal
f74a802d09 Test and fix #1919: Error resuming training 2018-02-02 02:32:40 +01:00
Matthew Honnibal
85c942a6e3 Dont overwrite pretrained_dims setting from cfg. Fixes #1727 2018-01-23 19:10:49 +01:00
Matthew Honnibal
fe4748fc38
Merge pull request #1870 from avadhpatel/master
Model Load Performance Improvement by more than 5x
2018-01-22 00:05:15 +01:00
Avadh Patel
a517df55c8 Small fix
Signed-off-by: Avadh Patel <avadh4all@gmail.com>
2018-01-21 15:20:45 -06:00
Avadh Patel
5b5029890d Merge branch 'perfTuning' into perfTuningMaster
Signed-off-by: Avadh Patel <avadh4all@gmail.com>
2018-01-21 15:20:00 -06:00
Matthew Honnibal
203d2ea830 Allow multitask objectives to be added to the parser and NER more easily 2018-01-21 19:37:02 +01:00
Avadh Patel
75903949da Updated model building after suggestion from Matthew
Signed-off-by: Avadh Patel <avadh4all@gmail.com>
2018-01-18 06:51:57 -06:00
Avadh Patel
fe879da2a1 Do not train model if its going to be loaded from disk
This saves significant time in loading a model from disk.

Signed-off-by: Avadh Patel <avadh4all@gmail.com>
2018-01-17 06:16:07 -06:00
Avadh Patel
2146faffee Do not train model if its going to be loaded from disk
This saves significant time in loading a model from disk.

Signed-off-by: Avadh Patel <avadh4all@gmail.com>
2018-01-17 06:04:22 -06:00
Matthew Honnibal
f29c3925ee Fix more efficient nonproj 2017-11-23 12:48:00 +00:00
Matthew Honnibal
db5c714ad2 Improve efficiency of deprojectivization 2017-11-23 12:31:34 +00:00
Matthew Honnibal
d274d3a3b9 Let beam forward use minibatches 2017-11-15 00:51:42 +01:00
Matthew Honnibal
855872f872 Remove state hashing 2017-11-14 23:36:46 +01:00
Matthew Honnibal
2512ea9eeb Fix memory leak in beam parser 2017-11-14 02:11:40 +01:00
Matthew Honnibal
ca73d0d8fe Cleanup states after beam parsing, explicitly 2017-11-13 18:18:26 +01:00
Matthew Honnibal
63ef9a2e73 Remove __dealloc__ from ParserBeam 2017-11-13 18:18:08 +01:00
Matthew Honnibal
25859dbb48 Return optimizer from begin_training, creating if necessary 2017-11-06 14:26:49 +01:00
Matthew Honnibal
2b35bb76ad Fix tensorizer on GPU 2017-11-05 15:34:40 +01:00
Matthew Honnibal
3ca16ddbd4 Merge branch 'develop' of https://github.com/explosion/spaCy into develop 2017-11-04 00:25:02 +01:00
Matthew Honnibal
98c29b7912 Add padding vector in parser, to make gradient more correct 2017-11-04 00:23:23 +01:00
Matthew Honnibal
13c8881d2f Expose parser's tok2vec model component 2017-11-03 20:20:59 +01:00
Matthew Honnibal
7fea845374 Remove print statement 2017-11-03 14:04:51 +01:00
Matthew Honnibal
a5b05f85f0 Set Doc.tensor attribute in parser 2017-11-03 11:21:00 +01:00
Matthew Honnibal
7698903617 Fix GPU usage 2017-10-31 02:33:16 +01:00
Matthew Honnibal
a0c7dabb72 Fix bug in 8-token parser features 2017-10-28 23:01:35 +00:00
Matthew Honnibal
b713d10d97 Switch to 13 features in parser 2017-10-28 23:01:14 +00:00
Matthew Honnibal
5414e2f14b Use missing features in parser 2017-10-28 16:45:54 +00:00
Matthew Honnibal
64e4ff7c4b Merge 'tidy-up' changes into branch. Resolve conflicts 2017-10-28 13:16:06 +02:00
Explosion Bot
b22e42af7f Merge changes to parser and _ml 2017-10-28 11:52:10 +02:00
ines
b4d226a3f1 Tidy up syntax 2017-10-27 19:45:57 +02:00
ines
9c89e2cdef Remove unused syntax iterators (now in language data) 2017-10-27 18:09:53 +02:00
ines
e33b7e0b3c Tidy up parser and ML 2017-10-27 14:39:30 +02:00
Matthew Honnibal
531142a933 Merge remote-tracking branch 'origin/develop' into feature/better-parser 2017-10-27 12:34:48 +00:00
Matthew Honnibal
75a637fa43 Remove redundant imports from _ml 2017-10-27 10:19:56 +00:00
Matthew Honnibal
bb25bdcd92 Adjust call to scatter_add for the new version 2017-10-27 01:16:55 +00:00
Matthew Honnibal
90d1d9b230 Remove obsolete parser code 2017-10-26 13:22:45 +02:00
Matthew Honnibal
33f8c58782 Remove obsolete parser.pyx 2017-10-26 12:42:05 +02:00
Matthew Honnibal
35977bdbb9 Update better-parser branch with develop 2017-10-26 00:55:53 +00:00
ines
18aae423fb Remove import of non-existing function 2017-10-25 15:54:10 +02:00
ines
5117a7d24d Fix whitespace 2017-10-25 15:54:02 +02:00
Matthew Honnibal
075e8118ea Update from develop 2017-10-25 12:45:21 +02:00
Matthew Honnibal
dd5b2d8fa3 Check for out-of-memory when calling calloc. Closes #1446 2017-10-24 12:40:47 +02:00
Matthew Honnibal
e7556ff048 Fix non-maxout parser 2017-10-23 18:16:23 +02:00
Matthew Honnibal
f111b228e0 Fix re-parsing of previously parsed text
If a Doc object had been previously parsed, it was possible for
invalid parses to be added. There were two problems:

1) The parse was only being partially erased
2) The RightArc action was able to create a 1-cycle.

This patch fixes both errors, and avoids resetting the parse if one is
present. In theory this might allow a better parse to be predicted by
running the parser twice.

Closes #1253.
2017-10-20 16:27:36 +02:00
Matthew Honnibal
1036798155 Make parser consistent if maxout==1 2017-10-20 16:24:16 +02:00
Matthew Honnibal
827cd8a883 Fix support of maxout pieces in parser 2017-10-20 03:07:17 +02:00
Matthew Honnibal
a8850b4282 Remove redundant PrecomputableMaxouts class 2017-10-19 20:27:34 +02:00
Matthew Honnibal
b00d0a2c97 Fix bias in parser 2017-10-19 18:42:11 +02:00
Matthew Honnibal
b54b4b8a97 Make parser_maxout_pieces hyper-param work 2017-10-19 13:45:18 +02:00
Matthew Honnibal
15e5a04a8d Clean up more depth=0 conditional code 2017-10-19 01:48:43 +02:00
Matthew Honnibal
906c50ac59 Fix loop typing, that caused error on windows 2017-10-19 01:48:39 +02:00
Matthew Honnibal
960788aaa2 Eliminate dead code in parser, and raise errors for obsolete options 2017-10-19 00:42:34 +02:00
Matthew Honnibal
bbfd7d8d5d Clean up parser multi-threading 2017-10-19 00:25:21 +02:00
Matthew Honnibal
f018f2030c Try optimized parser forward loop 2017-10-18 21:48:00 +02:00
Matthew Honnibal
633a75c7e0 Break parser batches into sub-batches, sorted by length. 2017-10-18 21:45:01 +02:00
Matthew Honnibal
908f44c3fe Disable history features by default 2017-10-12 14:56:11 +02:00
Matthew Honnibal
cecfcc7711 Set default hyper params back to 'slow' settings 2017-10-12 13:12:26 +02:00
Matthew Honnibal
807e109f2b Merge branch 'develop' of https://github.com/explosion/spaCy into develop 2017-10-11 02:47:59 -05:00
Matthew Honnibal
6e552c9d83 Prune number of non-projective labels more aggressiely 2017-10-11 02:46:44 -05:00
Matthew Honnibal
188f620046 Improve parser defaults 2017-10-11 09:43:48 +02:00
Matthew Honnibal
3065f12ef2 Make add parser label work for hidden_depth=0 2017-10-10 22:57:31 +02:00
Matthew Honnibal
8265b90c83 Update parser defaults 2017-10-09 21:55:20 -05:00
Matthew Honnibal
09d61ada5e Merge pull request #1396 from explosion/feature/pipeline-management
💫 Improve pipeline and factory management
2017-10-10 04:29:54 +02:00
Matthew Honnibal
d8a2506023 Merge pull request #1401 from explosion/feature/add-parser-action
💫 Allow labels to be added to pre-trained parser and NER modes
2017-10-09 04:57:51 +02:00
Matthew Honnibal
d43a83e37a Allow parser.add_label for pretrained models 2017-10-09 03:35:40 +02:00
Matthew Honnibal
4cc84b0234 Prohibit Break when sent_start < 0 2017-10-09 00:02:45 +02:00
Matthew Honnibal
e938bce320 Adjust parsing transition system to allow preset sentence segments. 2017-10-08 23:53:34 +02:00
Matthew Honnibal
20309fb9db Make history features default to zero 2017-10-08 20:32:14 +02:00
Matthew Honnibal
42b401d08b Change default hidden depth to 1 2017-10-07 21:05:21 -05:00
Matthew Honnibal
92c5d78b42 Unhack NER.add_action 2017-10-07 19:02:40 +02:00
Matthew Honnibal
3d22ccf495 Update default hyper-parameters 2017-10-07 07:16:41 -05:00
Matthew Honnibal
0384f08218 Trigger nonproj.deprojectivize as a postprocess 2017-10-07 02:00:47 +02:00
Matthew Honnibal
8be46d766e Remove print statement 2017-10-06 16:19:02 -05:00
Matthew Honnibal
8e731009fe Fix parser config serialization 2017-10-06 13:50:52 -05:00
Matthew Honnibal
16ba6aa8a6 Fix parser config serialization 2017-10-06 13:17:31 -05:00
Matthew Honnibal
c66399d8ae Fix depth definition with history features 2017-10-06 06:20:05 -05:00
Matthew Honnibal
5c750a9c2f Reserve 0 for 'missing' in history features 2017-10-06 06:10:13 -05:00
Matthew Honnibal
21d11936fe Fix significant train/test skew error in history feats 2017-10-06 06:08:50 -05:00
Matthew Honnibal
555d8c8bff Fix beam history features 2017-10-05 22:21:50 -05:00
Matthew Honnibal
b0618def8d Add support for 2-token state option 2017-10-05 21:54:12 -05:00
Matthew Honnibal
363aa47b40 Clean up dead parsing code 2017-10-05 21:53:49 -05:00
Matthew Honnibal
ca12764772 Enable history features for beam parser 2017-10-05 21:53:29 -05:00
Matthew Honnibal
e25ffcb11f Move history size under feature flags 2017-10-05 19:38:13 -05:00
Matthew Honnibal
943af4423a Make depth setting in parser work again 2017-10-04 20:06:05 -05:00
Matthew Honnibal
246612cb53 Merge remote-tracking branch 'origin/develop' into feature/parser-history-model 2017-10-03 16:56:42 -05:00
Matthew Honnibal
5454b20cd7 Update thinc imports for 6.9 2017-10-03 20:07:17 +02:00
Matthew Honnibal
4a59f6358c Fix thinc imports 2017-10-03 19:21:26 +02:00
Matthew Honnibal
dc3c791947 Fix history size option 2017-10-03 13:41:23 +02:00
Matthew Honnibal
278a4c17c6 Fix history features 2017-10-03 13:27:10 +02:00
Matthew Honnibal
b50a359e11 Add support for history features in parsing models 2017-10-03 12:44:01 +02:00
Matthew Honnibal
ee41e4fea7 Support history features in stateclass 2017-10-03 12:43:48 +02:00
Matthew Honnibal
cdb2d83e16 Pass dropout in parser 2017-09-28 18:47:13 -05:00
Matthew Honnibal
158e177cae Fix default embed size 2017-09-28 08:25:23 -05:00
Matthew Honnibal
1a37a2c0a0 Update training defaults 2017-09-27 11:48:07 -05:00
Matthew Honnibal
3274b46a0d Try to fix compile error on Windows 2017-09-26 09:05:53 -05:00
Matthew Honnibal
5056743ad5 Fix parser serialization 2017-09-26 06:44:56 -05:00
Matthew Honnibal
bf917225ab Allow multi-task objectives during training 2017-09-26 05:42:52 -05:00
Matthew Honnibal
4348c479fc Merge pre-trained vectors and noshare patches 2017-09-22 20:07:28 -05:00
Matthew Honnibal
0795857dcb Fix beam parsing 2017-09-23 02:59:53 +02:00
Matthew Honnibal
d9124f1aa3 Add link_vectors_to_models function 2017-09-22 09:38:22 -05:00
Matthew Honnibal
20193371f5 Don't share CNN, to reduce complexities 2017-09-21 14:59:48 +02:00
Matthew Honnibal
24e85c2048 Pass values for CNN maxout pieces option 2017-09-20 19:16:12 -05:00
Matthew Honnibal
2489dcaccf Fix serialization of parser 2017-09-19 23:42:12 +02:00
Matthew Honnibal
2b0efc77ae Fix wiring of pre-trained vectors in parser loading 2017-09-17 05:47:34 -05:00
Matthew Honnibal
31c2e91c35 Fix wiring of pre-trained vectors in parser loading 2017-09-17 05:46:55 -05:00
Matthew Honnibal
c003c561c3 Revert NER action loading change, for model compatibility 2017-09-17 05:46:03 -05:00
Matthew Honnibal
43210abacc Resolve fine-tuning conflict 2017-09-17 05:30:04 -05:00
Matthew Honnibal
5ff2491f24 Pass option for pre-trained vectors in parser 2017-09-16 12:47:21 -05:00
Matthew Honnibal
8665a77f48 Fix feature error in NER 2017-09-16 12:46:57 -05:00
Matthew Honnibal
f730d07e4e Fix prange error for Windows 2017-09-16 00:25:33 +02:00
Matthew Honnibal
8b481e0465 Remove redundant brackets 2017-09-15 10:38:08 +02:00
Matthew Honnibal
8c503487af Fix lookup of missing NER actions 2017-09-14 16:59:45 +02:00
Matthew Honnibal
664c5af745 Revert padding in parser 2017-09-14 16:59:25 +02:00
Matthew Honnibal
c6395b057a Improve parser feature extraction, for missing values 2017-09-14 16:18:02 +02:00
Matthew Honnibal
daf869ab3b Fix add_action for NER, so labelled 'O' actions aren't added 2017-09-14 16:16:41 +02:00