* Refactor sense tagger to get rid of intermediary layers

spacy/sense_tagger.pyx

@@ -1,13 +1,17 @@
-from thinc.api cimport Example
-from thinc.typedefs cimport atom_t
-
 from .typedefs cimport flags_t
 from .structs cimport TokenC
 from .strings cimport StringStore
 from .tokens cimport Tokens
-from ._ml cimport Model
 from .senses cimport POS_SENSES, N_SENSES, encode_sense_strs
 from .gold cimport GoldParse
+from .parts_of_speech cimport NOUN, VERB
+
+from thinc.learner cimport LinearModel
+from thinc.features cimport Extractor
+
+from thinc.typedefs cimport atom_t, weight_t, feat_t
+
+from os import path
 
 
 
@@ -173,6 +177,8 @@ cdef int fill_token(atom_t* ctxt, const TokenC* token) except -1:
 
 
 cdef int fill_context(atom_t* ctxt, const TokenC* token) except -1:
+    # NB: we have padding to keep us safe here
+    # See tokens.pyx
     fill_token(&ctxt[P2W], token - 2)
     fill_token(&ctxt[P1W], token - 1)
 
@@ -185,62 +191,79 @@ cdef int fill_context(atom_t* ctxt, const TokenC* token) except -1:
 
 cdef class SenseTagger:
     cdef readonly StringStore strings
-    cdef readonly Model model
+    cdef readonly LinearModel model
+    cdef readonly Extractor extractor
+    cdef readonly model_dir
 
     def __init__(self, StringStore strings, model_dir):
-        self.strings = strings
+        if model_dir is not None and path.isdir(model_dir):
+            model_dir = path.join(model_dir, 'model')
+
         templates = unigrams + bigrams + trigrams
-        self.model = Model(N_SENSES, templates, model_dir)
+        self.extractor = Extractor(templates)
+        self.model = LinearModel(N_SENSES, self.extractor.n_templ)
+        self.model_dir = model_dir
+        if self.model_dir and path.exists(self.model_dir):
+            self.model.load(self.model_dir, freq_thresh=0)
+        self.strings = strings
 
     def __call__(self, Tokens tokens):
-        eg = Example(self.model.n_classes, CONTEXT_SIZE, self.model.n_feats,
-                     self.model.n_feats)
-        cdef int i
+        cdef atom_t[CONTEXT_SIZE] context
+        cdef int i, guess, n_feats
+        cdef const TokenC* token
         for i in range(tokens.length):
-            n_valid = self._set_valid(<bint*>eg.c.is_valid, pos_senses(&tokens.data[i]))
-            if n_valid >= 1:
-                fill_context(eg.c.atoms, &tokens.data[i])
-                self.model.predict(eg)
-                tokens.data[i].sense = eg.c.guess
+            token = &tokens.data[i]
+            if token.pos in (NOUN, VERB):
+                fill_context(context, token)
+                feats = self.extractor.get_feats(context, &n_feats)
+                scores = self.model.get_scores(feats, n_feats)
+                tokens.data[i].sense = self.best_in_set(scores, POS_SENSES[<int>token.pos])
 
     def train(self, Tokens tokens, GoldParse gold):
-        eg = Example(self.model.n_classes, CONTEXT_SIZE, self.model.n_feats+1,
-                     self.model.n_feats+1)
-        cdef int i
+        cdef int i, j
         for i, ssenses in enumerate(gold.ssenses):
             if ssenses:
                 gold.c.ssenses[i] = encode_sense_strs(ssenses)
             else:
                 gold.c.ssenses[i] = pos_senses(&tokens.data[i])
+        
+        cdef atom_t[CONTEXT_SIZE] context
+        cdef int n_feats
+        cdef feat_t f_key
+        cdef int f_i
         cdef int cost = 0
         for i in range(tokens.length):
-            if tokens.data[i].lex.senses == 0 or tokens.data[i].lex.senses == 1:
-                continue
-            self._set_costs(<bint*>eg.c.is_valid, eg.c.costs, gold.c.ssenses[i])
-            fill_context(eg.c.atoms, &tokens.data[i])
-
-            self.model.train(eg)
-
-            tokens.data[i].sense = eg.c.guess
-            cost += eg.c.cost
+            token = &tokens.data[i]
+            if token.pos in (NOUN, VERB) \
+            and token.lex.senses >= 2 \
+            and gold.c.ssenses[i] >= 2:
+                fill_context(context, token)
+                feats = self.extractor.get_feats(context, &n_feats)
+                scores = self.model.get_scores(feats, n_feats)
+                token.sense = self.best_in_set(scores, POS_SENSES[<int>token.pos])
+                best = self.best_in_set(scores, gold.c.ssenses[i])
+                guess_counts = {}
+                gold_counts = {}
+                if token.sense != best:
+                    for j in range(n_feats):
+                        f_key = feats[j].key
+                        f_i = feats[j].i
+                        feat = (f_i, f_key)
+                        gold_counts[feat]  = gold_counts.get(feat, 0) + 1.0
+                        guess_counts[feat] = guess_counts.get(feat, 0) - 1.0
+                #self.model.update({token.sense: guess_counts, best: gold_counts})
         return cost
 
-    cdef int _set_valid(self, bint* is_valid, flags_t senses) except -1:
-        cdef int n_valid
-        cdef flags_t bit
-        is_valid[0] = False
-        for bit in range(1, N_SENSES):
-            is_valid[bit] = senses & (1 << bit)
-            n_valid += is_valid[bit]
-        return n_valid
-
-    cdef int _set_costs(self, bint* is_valid, int* costs, flags_t senses):
-        cdef flags_t bit
-        is_valid[0] = False
-        costs[0] = 1
-        for bit in range(1, N_SENSES):
-            is_valid[bit] = True 
-            costs[bit] = 0 if (senses & (1 << bit)) else 1
+    cdef int best_in_set(self, const weight_t* scores, flags_t senses) except -1:
+        cdef weight_t max_ = 0
+        cdef int argmax = -1
+        cdef flags_t i
+        for i in range(N_SENSES):
+            if (senses & (1 << i)) and (argmax == -1 or scores[i] > max_):
+                max_ = scores[i]
+                argmax = i
+        assert argmax >= 0
+        return argmax
 
 
 cdef flags_t pos_senses(const TokenC* token) nogil: