Rooba
/
spaCy
mirror of https://github.com/explosion/spaCy.git
1
1
Fork 0
Code Issues Packages Projects Releases Wiki Activity
8,688 Commits 70 Branches 177 Tags 298 MiB
Commit Graph

58 Commits

Author SHA1 Message Date
Matthew Honnibal 2512ea9eeb Fix memory leak in beam parser 2017-11-14 02:11:40 +01:00
ines b4d226a3f1 Tidy up syntax 2017-10-27 19:45:57 +02:00
Matthew Honnibal 363aa47b40 Clean up dead parsing code 2017-10-05 21:53:49 -05:00
Matthew Honnibal ee41e4fea7 Support history features in stateclass 2017-10-03 12:43:48 +02:00
Matthew Honnibal 3d5a536eaa Improve efficiency of parser batching 2017-05-26 11:31:23 -05:00
Matthew Honnibal a9edb3aa1d Improve integration of NN parser, to support unified training API 2017-05-15 21:53:27 +02:00
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 6782eedf9b Tmp GPU code 2017-05-07 11:04:24 -05:00
Matthew Honnibal f99f5b75dc working residual net 2017-05-07 03:57:26 +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
ines 0739ae7b76 Tidy up and fix formatting and imports 2017-04-15 13:05:15 +02:00
Matthew Honnibal 274a4d4272 Fix queue Python property in StateClass 2016-10-16 17:04:41 +02:00
Wolfgang Seeker d99a9cbce9 different handling of space tokens 
space tokens are now always attached to the previous non-space token
there are two exceptions:
leading space tokens are attached to the first following non-space token
in input that consists exclusively of space tokens, the last space token
is the head of all others.
 2016-04-13 15:28:28 +02:00
Matthew Honnibal daaad66448 * Now fully proxied 2016-02-01 02:37:08 +01:00
Matthew Honnibal 7a0e3bb9c1 * Continue proxying. Some problem currently 2016-02-01 02:22:21 +01:00
Matthew Honnibal 2169bbb7ea * Shadow StateClass with StateC, to start proxying 2016-02-01 01:16:14 +01:00
Matthew Honnibal 04177debd0 * Unwind limit to sentence boundary detection that prevents it from inserting boundaries on whitespace. Replace it with a check for whitespace in StateClass.fast_forward, so that whitespace is LeftArced when it's on the stack. This should prevent the previous problem of whitespace-only sentences. Should fix Issue #184, but may cause further problems. Needs testing. 2016-01-19 02:54:15 +01:00
Matthew Honnibal a06e3c8963 * Fix bone-headed mistake in StateClass.E 2015-11-07 07:35:28 +11:00
Matthew Honnibal d24b8509e4 * Correct screw ups from the previous commits 2015-11-07 06:51:41 +11:00
Matthew Honnibal 9285f01d26 * Fix broken StateClass.E tracking 2015-11-07 06:06:39 +11:00
Matthew Honnibal 2733816b7b * Fix whitespace 2015-11-07 05:31:06 +11:00
Matthew Honnibal b65633f270 * Fix function that returns nth entity in StateClass. Was only returning the first. 2015-11-07 05:29:11 +11:00
Matthew Honnibal 0e24d099a1 * Fix L/R edge bug, by ensuring l_edge and r_edge are preset, and fixing the way the edge update in del_arc. Bugs keep arising here because the edges are absolute positions, where everything else is relative. I'm also not 100% convinced that del_arc is handled correctly. Do we need to update the parents? 2015-09-09 03:40:44 +02:00
Matthew Honnibal 0653288fa5 * Fix stateclass.queue 2015-08-09 00:39:02 +02:00
Matthew Honnibal 7bafc789e7 * Add stack and queue properties to stateclass, for python access 2015-08-08 23:32:42 +02:00
Matthew Honnibal 823ef4a00b * Remove profile declarations 2015-07-25 18:13:06 +02:00
Matthew Honnibal 55f1042443 * Improve efficiency of L and R features, correcting the non-linear-in-length problem. 2015-07-09 12:17:26 +02:00
Matthew Honnibal 70d2acb579 * Fix edge features 2015-07-09 12:15:01 +02:00
Matthew Honnibal 3bb5876c5a * Inline methods in StateClass 2015-06-29 01:10:14 +02:00
Matthew Honnibal a02fd3af5d * Check valency in L and R feature methods, to make feaure calculation faster 2015-06-29 00:27:56 +02:00
Matthew Honnibal b06962f18b * Pad buffers in state 2015-06-28 10:36:14 +02:00
Matthew Honnibal d5b4090705 * Add profile directive 2015-06-28 06:19:33 +02:00
Matthew Honnibal 02b171ee67 * Bug fixes to edge calculation 2015-06-24 04:28:02 +02:00
Matthew Honnibal 7ebfe4b983 * Fixes to edge features 2015-06-23 16:32:54 +02:00
Matthew Honnibal 35c290bee4 * Fix edge features 2015-06-23 15:50:56 +02:00
Matthew Honnibal a7bf7b0626 * Rename sent_start to sent_end, to reflect its new usage in the Break transition 2015-06-23 05:39:43 +02:00
Matthew Honnibal ee3e56f27b * Fix bounds checking on entities 2015-06-23 04:35:08 +02:00
Matthew Honnibal 065c2e1d2d * Add some bounds checking around state arrays 2015-06-23 04:13:09 +02:00
Matthew Honnibal f01b3d043e * Add padding to arrays in stateclass. May be papering over a deeper bug. 2015-06-23 03:03:41 +02:00
Matthew Honnibal 9b13d11ab3 * Fix handling of entities in StateClass 2015-06-16 23:35:21 +02:00
Matthew Honnibal e0984ca139 * Fix valency features in StateClass 2015-06-14 17:50:26 +02:00
Matthew Honnibal 763cbd23d5 * Upd stateclass.print_state 2015-06-14 17:44:29 +02:00
Matthew Honnibal 15e177d7a1 * Fixes to unshift/fast-forward strategy. Getting 91.55 greedy on NW dev, gold preproc 2015-06-12 01:50:23 +02:00
Matthew Honnibal afd77a529b * Prepare for break transition, with fast-forwarding. 86.5 on 1k nw gold preproc 2015-06-10 14:08:30 +02:00
Matthew Honnibal 495f528709 * Add support for sentence breaks in stateclass 2015-06-10 12:34:28 +02:00
Matthew Honnibal bb09b5d91a * Fix shifted bit vector in stateclass --- should reflect whether the word has been *unshifted*. 2015-06-10 11:33:09 +02:00