Commit Graph

102 Commits

Author SHA1 Message Date
Matthew Honnibal
df8179ca4f * Add separate Param and AdadeltaParam classes. AdadeltaParam seems broken. 2015-06-29 16:39:16 +02:00
Matthew Honnibal
1dff04acb5 * Apply regularization to the softmax, not the bias 2015-06-29 11:45:38 +02:00
Matthew Honnibal
ca30fe1582 * Use He initialization trick 2015-06-29 10:56:02 +02:00
Matthew Honnibal
fc34e1b6e4 * Move Theano functions into nn_train.py script 2015-06-29 07:09:16 +02:00
Matthew Honnibal
fe7b24ecef * whitespace 2015-06-28 11:37:17 +02:00
Matthew Honnibal
7b8275fcc4 * Wire hyperparameters to script interface 2015-06-28 11:37:17 +02:00
Matthew Honnibal
897dd0dd0b * Merge changes, and adjust Example to use memoryview 2015-06-28 11:36:11 +02:00
Matthew Honnibal
ef97b90833 * Fix token scoring 2015-06-28 06:22:18 +02:00
Matthew Honnibal
34c0ef2ee8 * Don't compile the orig_arc_eager and tree_arc_eager modules used for the EMNLP paper 2015-06-23 05:38:17 +02:00
Matthew Honnibal
59e9f9153c * Remove projectivity constraint in train.py, but raise Exception if non-projective sentence is encountered, since we've told GoldParse to projectivize 2015-06-23 05:04:46 +02:00
Matthew Honnibal
839e5038b7 * Raise exception on non-projective input 2015-06-23 00:01:55 +02:00
Matthew Honnibal
4dad4058c3 * Uncomment NER training 2015-06-16 23:36:54 +02:00
Matthew Honnibal
5699585278 * Use tree_arc_eager system as baseline in experiments 2015-06-15 08:23:43 +02:00
Matthew Honnibal
4841f8ad5e * Set transition system early 2015-06-15 02:54:12 +02:00
Matthew Honnibal
bcfdf126a4 * Add toggle for OrigArcEager system 2015-06-14 20:28:14 +02:00
Matthew Honnibal
c500d72dc2 * Temporarily disable NER, and wire up the verbose flag during training 2015-06-14 17:45:31 +02:00
Matthew Honnibal
ac422492cf * Fix write_parses mode of bin/parser/train.py 2015-06-07 19:08:48 +02:00
Matthew Honnibal
4073533e28 * Upd munge_ewtb for the new json format 2015-06-06 02:10:33 +02:00
Matthew Honnibal
6a1341b29e * Add tb pre-process script 2015-06-06 01:59:44 +02:00
Matthew Honnibal
1736fc5a67 * Add more options to bin/parser/train 2015-06-05 23:49:26 +02:00
Matthew Honnibal
362f87dc3a * Update input corruption method to work with lists as well as trings 2015-06-05 19:33:32 +02:00
Matthew Honnibal
0aed9c9a33 * Fix train.py 2015-06-05 15:50:24 +02:00
Matthew Honnibal
8466600add * Clean up train.py, removing unused tag jackknifing code 2015-06-05 15:01:28 +02:00
Matthew Honnibal
e772b48dcd * Skip sentences of length 1 in training 2015-06-05 02:29:03 +02:00
Matthew Honnibal
e822df0867 * Fix bugs in new greedy/beam parser 2015-06-02 02:01:33 +02:00
Matthew Honnibal
70a7ad89ca * Removed unused imports from train.py 2015-06-02 00:59:09 +02:00
Matthew Honnibal
a3de20118e * Wire up beam-width command line argument 2015-06-02 00:54:12 +02:00
Matthew Honnibal
08044ea70c * Remove try/except around parser.train 2015-05-31 15:21:56 +02:00
Matthew Honnibal
c8a553fe91 * Fix cluster initialization 2015-05-31 15:21:28 +02:00
Matthew Honnibal
d7cc2338e7 * Fix bug in train.py 2015-05-31 06:49:06 +02:00
Matthew Honnibal
c037f80638 * Add case expansion to Brown clusters 2015-05-31 05:50:50 +02:00
Matthew Honnibal
5ab0f233a1 * Ensure words in Brown clusters make it into the vocab, even if they're not in our probs list 2015-05-31 05:46:16 +02:00
Matthew Honnibal
d42dda0372 * Shuffle docs before doing jackknife partition --- otherwise we'll not get the right genre mixes... 2015-05-31 01:25:02 +02:00
Matthew Honnibal
4d8d490547 * Exclude empty sentences in prepare_treebank 2015-05-31 01:12:46 +02:00
Matthew Honnibal
d512d20d81 * Allow parser to jackknife POS tags before training. 2015-05-31 01:11:11 +02:00
Matthew Honnibal
6bbdcc5db5 * Fix gold_preproc flag in train.py 2015-05-30 05:23:02 +02:00
Matthew Honnibal
76300bbb1b * Use updated JSON format, with sentences below paragraphs. Allows use of gold preprocessing flag. 2015-05-30 01:25:46 +02:00
Matthew Honnibal
2d11739f28 * Change data format of JSON corpus, putting sentences into lists with the paragraph 2015-05-30 01:25:00 +02:00
Matthew Honnibal
784e577f45 * Check NER length matches conll length in prepare_treebank 2015-05-29 03:54:06 +02:00
Matthew Honnibal
b76bbbd12c * Read json files recursively from a directory, instead of requiring a single .json file 2015-05-29 03:52:55 +02:00
Matthew Honnibal
ef67ef7a4c * Recomment in training in train.py 2015-05-28 22:40:26 +02:00
Matthew Honnibal
5eb64eeb11 * Print json treebank by genre, instead of by large file 2015-05-28 22:40:01 +02:00
Matthew Honnibal
f42dc1f7d8 * Fix evaluate method in train.py, to use sentences which don't have raw text 2015-05-28 16:30:23 +02:00
Matthew Honnibal
a7cee46fe9 * Update train.py, to support paragraphs where there's no raw_text 2015-05-27 19:14:02 +02:00
Matthew Honnibal
ef1333cf89 * Have prepare_treebank read train/dev/test IDs. 2015-05-27 17:35:05 +02:00
Matthew Honnibal
e140e03516 * Read in OntoNotes. Doesn't support train/test/dev split yet 2015-05-27 17:04:29 +02:00
Matthew Honnibal
895060e774 * Ensure tagger and NER are trained, even if non-projective problem 2015-05-27 03:16:21 +02:00
Matthew Honnibal
32ae2cdabe * In prepare_treebank, move ner into the token descriptions 2015-05-26 19:52:39 +02:00
Matthew Honnibal
61885aee76 * Work on prepare_treebank script, adding NER to it 2015-05-26 19:28:29 +02:00
Matthew Honnibal
15bbbf4901 * Remove cruft from train.py 2015-05-25 07:54:10 +02:00