Commit Graph

545 Commits

Author SHA1 Message Date
Matthew Honnibal
3ff8c35a79 Move to contiguous buffer for token_ids and d_vectors 2017-05-20 04:17:30 -05:00
Matthew Honnibal
8b04b0af9f Remove freqs from transition_system 2017-05-20 02:20:48 -05:00
Matthew Honnibal
a1ba20e2b1 Fix over-run on parse_batch 2017-05-19 18:57:30 -05:00
Matthew Honnibal
e84de028b5 Remove 'rebatch' op, and remove min-batch cap 2017-05-19 18:16:36 -05:00
Matthew Honnibal
c12ab47a56 Remove state argument in pipeline. Other changes 2017-05-19 13:26:36 -05:00
Matthew Honnibal
c2c825127a Fix use_params and pipe methods 2017-05-18 08:30:59 -05:00
Matthew Honnibal
fc8d3a112c Add util.env_opt support: Can set hyper params through environment variables. 2017-05-18 04:36:53 -05:00
Matthew Honnibal
d2626fdb45 Fix name error in nn parser 2017-05-18 04:31:01 -05:00
Matthew Honnibal
793430aa7a Get spaCy train command working with neural network
* Integrate models into pipeline
* Add basic serialization (maybe incorrect)
* Fix pickle on vocab
2017-05-17 12:04:50 +02:00
Matthew Honnibal
8cf097ca88 Redesign training to integrate NN components
* Obsolete .parser, .entity etc names in favour of .pipeline
* Components no longer create models on initialization
* Models created by loading method (from_disk(), from_bytes() etc), or
    .begin_training()
* Add .predict(), .set_annotations() methods in components
* Pass state through pipeline, to allow components to share information
    more flexibly.
2017-05-16 16:17:30 +02:00
Matthew Honnibal
5211645af3 Get data flowing through pipeline. Needs redesign 2017-05-16 11:21:59 +02:00
Matthew Honnibal
a9edb3aa1d Improve integration of NN parser, to support unified training API 2017-05-15 21:53:27 +02:00
Matthew Honnibal
4b9d69f428 Merge branch 'v2' into develop
* Move v2 parser into nn_parser.pyx
* New TokenVectorEncoder class in pipeline.pyx
* New spacy/_ml.py module

Currently the two parsers live side-by-side, until we figure out how to
organize them.
2017-05-14 01:10:23 +02:00
Matthew Honnibal
5cac951a16 Move new parser to nn_parser.pyx, and restore old parser, to make tests pass. 2017-05-14 00:55:01 +02:00
Matthew Honnibal
f8c02b4341 Remove cupy imports from parser, so it can work on CPU 2017-05-14 00:37:53 +02:00
Matthew Honnibal
e6d71e1778 Small fixes to parser 2017-05-13 17:19:04 -05:00
Matthew Honnibal
188c0f6949 Clean up unused import 2017-05-13 17:18:27 -05:00
Matthew Honnibal
f85c8464f7 Draft support of regression loss in parser 2017-05-13 17:17:27 -05:00
Matthew Honnibal
827b5af697 Update draft of parser neural network model
Model is good, but code is messy. Currently requires Chainer, which may cause the build to fail on machines without a GPU.

Outline of the model:

We first predict context-sensitive vectors for each word in the input:

(embed_lower | embed_prefix | embed_suffix | embed_shape)
>> Maxout(token_width)
>> convolution ** 4

This convolutional layer is shared between the tagger and the parser. This prevents the parser from needing tag features.
To boost the representation, we make a "super tag" with POS, morphology and dependency label. The tagger predicts this
by adding a softmax layer onto the convolutional layer --- so, we're teaching the convolutional layer to give us a
representation that's one affine transform from this informative lexical information. This is obviously good for the
parser (which backprops to the convolutions too).

The parser model makes a state vector by concatenating the vector representations for its context tokens. Current
results suggest few context tokens works well. Maybe this is a bug.

The current context tokens:

* S0, S1, S2: Top three words on the stack
* B0, B1: First two words of the buffer
* S0L1, S0L2: Leftmost and second leftmost children of S0
* S0R1, S0R2: Rightmost and second rightmost children of S0
* S1L1, S1L2, S1R2, S1R, B0L1, B0L2: Likewise for S1 and B0

This makes the state vector quite long: 13*T, where T is the token vector width (128 is working well). Fortunately,
there's a way to structure the computation to save some expense (and make it more GPU friendly).

The parser typically visits 2*N states for a sentence of length N (although it may visit more, if it back-tracks
with a non-monotonic transition). A naive implementation would require 2*N (B, 13*T) @ (13*T, H) matrix multiplications
for a batch of size B. We can instead perform one (B*N, T) @ (T, 13*H) multiplication, to pre-compute the hidden
weights for each positional feature wrt the words in the batch. (Note that our token vectors come from the CNN
-- so we can't play this trick over the vocabulary. That's how Stanford's NN parser works --- and why its model
is so big.)

This pre-computation strategy allows a nice compromise between GPU-friendliness and implementation simplicity.
The CNN and the wide lower layer are computed on the GPU, and then the precomputed hidden weights are moved
to the CPU, before we start the transition-based parsing process. This makes a lot of things much easier.
We don't have to worry about variable-length batch sizes, and we don't have to implement the dynamic oracle
in CUDA to train.

Currently the parser's loss function is multilabel log loss, as the dynamic oracle allows multiple states to
be 0 cost. This is defined as:

(exp(score) / Z) - (exp(score) / gZ)

Where gZ is the sum of the scores assigned to gold classes. I'm very interested in regressing on the cost directly,
but so far this isn't working well.

Machinery is in place for beam-search, which has been working well for the linear model. Beam search should benefit
greatly from the pre-computation trick.
2017-05-12 16:09:15 -05:00
Matthew Honnibal
b44f7e259c Clean up unused parser code 2017-05-08 15:42:04 +02:00
Matthew Honnibal
17efb1c001 Change width 2017-05-08 08:40:13 -05:00
Matthew Honnibal
bef89ef23d Mergery 2017-05-08 08:29:36 -05:00
Matthew Honnibal
50ddc9fc45 Fix infinite loop bug 2017-05-08 07:54:26 -05:00
Matthew Honnibal
a66a4a4d0f Replace einsums 2017-05-08 14:46:50 +02:00
Matthew Honnibal
8d2eab74da Use PretrainableMaxouts 2017-05-08 14:24:55 +02:00
Matthew Honnibal
2e2268a442 Precomputable hidden now working 2017-05-08 11:36:37 +02:00
Matthew Honnibal
10682d35ab Get pre-computed version working 2017-05-08 00:38:35 +02:00
Matthew Honnibal
35458987e8 Checkpoint -- nearly finished reimpl 2017-05-07 23:05:01 +02:00
Matthew Honnibal
4441866f55 Checkpoint -- nearly finished reimpl 2017-05-07 22:47:06 +02:00
Matthew Honnibal
6782eedf9b Tmp GPU code 2017-05-07 11:04:24 -05:00
Matthew Honnibal
e420e5a809 Tmp 2017-05-07 07:31:09 -05:00
Matthew Honnibal
700979fb3c CPU/GPU compat 2017-05-07 04:01:11 +02:00
Matthew Honnibal
f99f5b75dc working residual net 2017-05-07 03:57:26 +02:00
Matthew Honnibal
bdf2dba9fb WIP on refactor, with hidde pre-computing 2017-05-07 02:02:43 +02:00
Matthew Honnibal
b439e04f8d Learning smoothly 2017-05-06 20:38:12 +02:00
Matthew Honnibal
08bee76790 Learns things 2017-05-06 18:24:38 +02:00
Matthew Honnibal
bcf4cd0a5f Learns things 2017-05-06 17:37:36 +02:00
Matthew Honnibal
8e48b58cd6 Gradients look correct 2017-05-06 16:47:15 +02:00
Matthew Honnibal
7e04260d38 Data running through, likely errors in model 2017-05-06 14:22:20 +02:00
Matthew Honnibal
ef4fa594aa Draft of NN parser, to be tested 2017-05-05 19:20:39 +02:00
Matthew Honnibal
ccaf26206b Pseudocode for parser 2017-05-04 12:17:59 +02:00
Matthew Honnibal
2da16adcc2 Add dropout optin for parser and NER
Dropout can now be specified in the `Parser.update()` method via
the `drop` keyword argument, e.g.

    nlp.entity.update(doc, gold, drop=0.4)

This will randomly drop 40% of features, and multiply the value of the
others by 1. / 0.4. This may be useful for generalising from small data
sets.

This commit also patches the examples/training/train_new_entity_type.py
example, to use dropout and fix the output (previously it did not output
the learned entity).
2017-04-27 13:18:39 +02:00
Matthew Honnibal
d2436dc17b Update fix for Issue #999 2017-04-23 18:14:37 +02:00
Matthew Honnibal
60703cede5 Ensure noun chunks can't be nested. Closes #955 2017-04-23 17:56:39 +02:00
Matthew Honnibal
4eef200bab Persist the actions within spacy.parser.cfg 2017-04-20 17:02:44 +02:00
Matthew Honnibal
137b210bcf Restore use of FTRL training 2017-04-16 18:02:42 +02:00
Matthew Honnibal
45464d065e Remove print statement 2017-04-15 16:11:43 +02:00
Matthew Honnibal
c76cb8af35 Fix training for new labels 2017-04-15 16:11:26 +02:00
Matthew Honnibal
4884b2c113 Refix StepwiseState 2017-04-15 16:00:28 +02:00
Matthew Honnibal
1a98e48b8e Fix Stepwisestate' 2017-04-15 13:35:01 +02:00
ines
0739ae7b76 Tidy up and fix formatting and imports 2017-04-15 13:05:15 +02:00
Matthew Honnibal
354458484c WIP on add_label bug during NER training
Currently when a new label is introduced to NER during training,
it causes the labels to be read in in an unexpected order. This
invalidates the model.
2017-04-14 23:52:17 +02:00
Matthew Honnibal
49e2de900e Add costs property to StepwiseState, to show which moves are gold. 2017-04-10 11:37:04 +02:00
Matthew Honnibal
cc36c308f4 Fix noun_chunk rules around coordination
Closes #693.
2017-04-07 17:06:40 +02:00
Matthew Honnibal
1bb7b4ca71 Add comment 2017-03-31 13:59:19 +02:00
Matthew Honnibal
47a3ef06a6 Unhack deprojetivization, moving it into pipeline
Previously the deprojectivize() call was attached to the transition
system, and only called for German. Instead it should be a separate
process, called after the parser. This makes it available for any
language. Closes #898.
2017-03-31 12:31:50 +02:00
Matthew Honnibal
a9b1f23c7d Enable regression loss for parser 2017-03-26 09:26:30 -05:00
Matthew Honnibal
b487b8735a Decrease beam density, and fix Python 3 problem in beam 2017-03-20 12:56:05 +01:00
Matthew Honnibal
c90dc7ac29 Clean up state initiatisation in transition system 2017-03-16 11:59:11 -05:00
Matthew Honnibal
a46933a8fe Clean up FTRL parsing stuff. 2017-03-16 11:58:20 -05:00
Matthew Honnibal
2611ac2a89 Fix scorer bug for NER, related to ambiguity between missing annotations and misaligned tokens 2017-03-16 09:38:28 -05:00
Matthew Honnibal
3d0833c3df Fix off-by-1 in parse features fill_context 2017-03-15 19:55:35 -05:00
Matthew Honnibal
4ef68c413f Approximate cost in Break transition, to speed things up a bit. 2017-03-15 16:40:27 -05:00
Matthew Honnibal
8543db8a5b Use ftrl optimizer in parser 2017-03-15 11:56:37 -05:00
Matthew Honnibal
d719f8e77e Use nogil in parser, and set L1 to 0.0 by default 2017-03-15 09:31:01 -05:00
Matthew Honnibal
c61c501406 Update beam-parser to allow parser to maintain nogil 2017-03-15 09:30:22 -05:00
Matthew Honnibal
c79b3129e3 Fix setting of empty lexeme in initial parse state 2017-03-15 09:26:53 -05:00
Matthew Honnibal
6c4108c073 Add header for beam parser 2017-03-11 12:45:12 -06:00
Matthew Honnibal
931feb3360 Allow beam parsing for NER 2017-03-11 11:12:01 -06:00
Matthew Honnibal
ca9c8c57c0 Add iteration argument to parser.update 2017-03-11 07:00:47 -06:00
Matthew Honnibal
d59c6926c1 I think this fixes the segfault 2017-03-11 06:58:34 -06:00
Matthew Honnibal
318b9e32ff WIP on beam parser. Currently segfaults. 2017-03-11 06:19:52 -06:00
Matthew Honnibal
b0d80dc9ae Update name of 'train' function in BeamParser 2017-03-10 14:35:43 -06:00
Matthew Honnibal
d11f1a4ddf Record negative costs in non-monotonic arc eager oracle 2017-03-10 11:22:04 -06:00
Matthew Honnibal
ecf91a2dbb Support beam parser 2017-03-10 11:21:21 -06:00
Matthew Honnibal
c62da02344 Use ftrl training, to learn compressed model. 2017-03-09 18:43:21 -06:00
Matthew Honnibal
40703988bc Use FTRL training in parser 2017-03-08 01:38:51 +01:00
Roman Inflianskas
66e1109b53 Add support for Universal Dependencies v2.0 2017-03-03 13:17:34 +01:00
Matthew Honnibal
97a1286129 Revert changes to tagger and parser for thinc 6 2017-01-09 10:08:34 -06:00
Matthew Honnibal
af81ac8bb0 Use thinc 6.0 2016-12-29 11:58:42 +01:00
Matthew Honnibal
bc0a202c9c Fix unicode problem in nonproj module 2016-11-25 17:29:17 -06:00
Matthew Honnibal
159e8c46e1 Merge old training fixes with newer state 2016-11-25 09:16:36 -06:00
Matthew Honnibal
39341598bb Fix NER label calculation 2016-11-25 09:02:22 -06:00
Matthew Honnibal
ca773a1f53 Tweak arc_eager n_gold to deal with negative costs, and improve error message. 2016-11-25 09:01:52 -06:00
Matthew Honnibal
608d8f5421 Pass cfg through parser, and have is_valid default to 1, not 0 when resetting state 2016-11-25 09:00:21 -06:00
Matthew Honnibal
b8c4f5ea76 Allow German noun chunks to work on Span
Update the German noun chunks iterator, so that it also works on Span objects.
2016-11-24 23:30:15 +11:00
Pokey Rule
3e3bda142d Add noun_chunks to Span 2016-11-24 10:47:20 +00:00
Matthew Honnibal
b86f8af0c1 Fix doc strings 2016-11-01 12:25:36 +01:00
Matthew Honnibal
708ea22208 Infer types in transition_system.pyx 2016-10-27 18:08:13 +02:00
Matthew Honnibal
301f3cc898 Fix Issue #429. Add an initialize_state method to the named entity recogniser that adds missing entity types. This is a messy place to add this, because it's strange to have the method mutate state. A better home for this logic could be found. 2016-10-27 18:01:55 +02:00
Matthew Honnibal
03a520ec4f Change signature of Parser.parseC, so that nr_class is read from the transition system. This allows the transition system to modify the number of actions in initialize_state. 2016-10-27 17:58:56 +02:00
Matthew Honnibal
a209b10579 Improve error message when oracle fails for non-projective trees, re Issue #571. 2016-10-24 20:31:30 +02:00
Matthew Honnibal
3e688e6d4b Fix issue #514 -- serializer fails when new entity type has been added. The fix here is quite ugly. It's best to add the entities ASAP after loading the NLP pipeline, to mitigate the brittleness. 2016-10-23 17:45:44 +02:00
Matthew Honnibal
59038f7efa Restore support for prior data format -- specifically, the labels field of the config. 2016-10-17 00:53:26 +02:00
Matthew Honnibal
7887ab3b36 Fix default use of feature_templates in parser 2016-10-16 21:41:56 +02:00
Matthew Honnibal
f787cd29fe Refactor the pipeline classes to make them more consistent, and remove the redundant blank() constructor. 2016-10-16 21:34:57 +02:00
Matthew Honnibal
274a4d4272 Fix queue Python property in StateClass 2016-10-16 17:04:41 +02:00
Matthew Honnibal
e8c8aa08ce Make action_name optional in StepwiseState 2016-10-16 17:04:16 +02:00
Matthew Honnibal
4fc56d4a31 Rename 'labels' to 'actions' in parser options 2016-10-16 11:42:26 +02:00
Matthew Honnibal
3259a63779 Whitespace 2016-10-16 01:47:28 +02:00