diff --git a/spacy/tagger.pyx b/spacy/tagger.pyx index 9ca15f4ad..df4957b4a 100644 --- a/spacy/tagger.pyx +++ b/spacy/tagger.pyx @@ -1,4 +1,5 @@ from __future__ import unicode_literals +cimport cython import json from os import path from collections import defaultdict @@ -10,6 +11,7 @@ from thinc.extra.eg cimport Example from thinc.structs cimport ExampleC from thinc.linear.avgtron cimport AveragedPerceptron from thinc.linalg cimport VecVec +from thinc.structs cimport FeatureC from .typedefs cimport attr_t from .tokens.doc cimport Doc @@ -125,14 +127,29 @@ cdef class TaggerNeuralNet(NeuralNet): cdef class CharacterTagger(NeuralNet): def __init__(self, n_classes, - depth=4, hidden_width=50, - chars_width=10, - words_width=20, shape_width=5, suffix_width=5, tags_width=20, - learn_rate=0.1): - input_length = 5 * chars_width * self.chars_per_word + 2 * tags_width - widths = [input_length] + [hidden_width] * depth + [n_classes] + depth=3, hidden_width=100, chars_width=5, tags_width=10, learn_rate=0.1, + left_window=2, right_window=2, tags_window=10, chars_per_word=8): + self.chars_per_word = chars_per_word + self.chars_width = chars_width + self.tags_width = tags_width + self.left_window = left_window + self.right_window = right_window + self.tags_window = tags_window + self.depth = depth + self.hidden_width = hidden_width + + input_length = self.left_window * self.chars_width * self.chars_per_word \ + + self.right_window * self.chars_width * self.chars_per_word \ + + 1 * self.chars_width * self.chars_per_word \ + + self.tags_window * self.tags_width + + widths = [input_length] + [self.hidden_width] * self.depth + [n_classes] + vector_widths = [chars_width, tags_width] - slots = [0] * 5 * self.chars_per_word + [1] * 2 + slots = [0] * self.left_window * self.chars_per_word \ + + [0] * self.right_window * self.chars_per_word \ + + [0] * 1 * self.chars_per_word \ + + [1] * self.tags_window NeuralNet.__init__( self, widths, @@ -141,36 +158,30 @@ cdef class CharacterTagger(NeuralNet): rho=1e-6, update_step='sgd') - cdef void set_featuresC(self, ExampleC* eg, const TokenC* tokens, object strings, int i) except *: - oov = '_' * self.chars_per_word + cdef void set_featuresC(self, ExampleC* eg, + const TokenC* tokens, object strings, const int i) except *: + cdef unicode oov = '' p2 = strings[i-2] if i >= 2 else oov p1 = strings[i-1] if i >= 1 else oov w = strings[i] n1 = strings[i+1] if (i+1) < len(strings) else oov n2 = strings[i+2] if (i+2) < len(strings) else oov - cdef int p = 0 - cdef int c cdef int chars_per_word = self.chars_per_word - cdef unicode string + eg.nr_feat = 0 for string in (p2, p1, w, n1, n2): - for c in range(chars_per_word / 2): - eg.features[p].i = p - eg.features[p].key = ord(string[c]) - eg.features[p].value = 1.0 if string[c] != u'_' else 0.0 - p += 1 - eg.features[p].i = p - eg.features[p].key = ord(string[-(c+1)]) - eg.features[p].value = 1.0 if string[-(c+1)] != u'_' else 0.0 - p += 1 - eg.features[p].key = tokens[i-1].tag - eg.features[p].value = 1.0 - eg.features[p].i = p - p += 1 - eg.features[p].key = tokens[i-2].tag - eg.features[p].value = 1.0 - eg.features[p].i = p - eg.nr_feat = p+1 - + set_character_features(&eg.features[eg.nr_feat], + string, chars_per_word) + eg.nr_feat += chars_per_word + cdef int hist + for hist in range(1, self.tags_window+1): + tag = tokens[i-hist].tag if hist <= i else 0 + eg.features[eg.nr_feat].key = tag + eg.features[eg.nr_feat].value = 1.0 + eg.nr_feat += 1 + cdef int p + for p in range(eg.nr_feat): + eg.features[p].i = p + def end_training(self): pass @@ -179,24 +190,9 @@ cdef class CharacterTagger(NeuralNet): property nr_feat: def __get__(self): - return self.chars_per_word * 5 + 2 + nr_word = self.left_window + self.right_window + 1 + return self.chars_per_word * nr_word + self.tags_window - property chars_per_word: - def __get__(self): - return 16 - - -def _pad(word, nr_char): - if len(word) == nr_char: - pass - elif len(word) > nr_char: - split = nr_char / 2 - word = word[:split] + word[-split:] - else: - word = word.ljust(nr_char, '_') - assert len(word) == nr_char, repr(word) - return word - cdef inline void _fill_from_token(atom_t* context, const TokenC* t) nogil: context[0] = t.lex.lower @@ -293,9 +289,6 @@ cdef class Tagger: def tag_names(self): return self.vocab.morphology.tag_names - def __reduce__(self): - return (self.__class__, (self.vocab, self.model), None, None) - def tag_from_strings(self, Doc tokens, object tag_strs): cdef int i for i in range(tokens.length): @@ -319,14 +312,13 @@ cdef class Tagger: widths=self.model.widths, nr_class=self.vocab.morphology.n_tags, nr_feat=self.model.nr_feat) - strings = [_pad(tok.text, self.model.chars_per_word) for tok in tokens] + strings = [tok.text for tok in tokens] for i in range(tokens.length): eg.reset() if tokens.c[i].pos == 0: self.model.set_featuresC(&eg.c, tokens.c, strings, i) self.model.predict_example(eg) - #self.model.set_scoresC(eg.c.scores, - # eg.c.features, eg.c.nr_feat) + guess = VecVec.arg_max_if_true(eg.c.scores, eg.c.is_valid, eg.c.nr_class) self.vocab.morphology.assign_tag(&tokens.c[i], guess) tokens.is_tagged = True @@ -345,7 +337,7 @@ cdef class Tagger: "gold tags, to maintain coarse-grained mapping.") raise ValueError(msg % tag) golds = [self.tag_names.index(g) if g is not None else -1 for g in gold_tag_strs] - strings = [_pad(tok.text, self.model.chars_per_word) for tok in tokens] + strings = [tok.text for tok in tokens] cdef int correct = 0 cdef Pool mem = Pool() cdef Example eg = Example( @@ -357,16 +349,35 @@ cdef class Tagger: eg.reset() self.model.set_featuresC(&eg.c, tokens.c, strings, i) eg.costs = [golds[i] not in (j, -1) for j in range(eg.c.nr_class)] + self.model.train_example(eg) - - #self.model.set_scoresC(eg.c.scores, - # eg.c.features, eg.c.nr_feat) - # - #self.model.updateC(&eg.c) - + self.vocab.morphology.assign_tag(&tokens.c[i], eg.guess) + correct += eg.cost == 0 self.freqs[TAG][tokens.c[i].tag] += 1 tokens.is_tagged = True tokens._py_tokens = [None] * tokens.length return correct + + +@cython.cdivision(True) +cdef int set_character_features(FeatureC* feat, unicode string, int chars_per_word) except -1: + cdef unicode oov = '' + cdef int chars_per_side = min(chars_per_word / 2, len(string)) + # Fill from start + cdef int c + for c in range(chars_per_side): + feat.key = ord(string[c]) + feat.value = 1.0 + feat += 1 + # If word is too short, zero this part of the array + for c in range(chars_per_side, chars_per_word - chars_per_side): + feat.key = 0 + feat.value = 0 + feat += 1 + # Fill suffix + for c in range(chars_per_side): + feat.key = ord(string[-(c+1)]) + feat.value = 1.0 + feat += 1