pystruct · vene · Aug 14, 2013 · Aug 14, 2013
diff --git a/examples/syllable_chain.py b/examples/syllable_chain.py
@@ -0,0 +1,87 @@
+from __future__ import print_function
+
+try:
+    from urllib.request import urlopen  # Py3k
+except ImportError:
+    from urllib2 import urlopen  # Py2
+
+import numpy as np
+
+from sklearn.feature_extraction import FeatureHasher
+from sklearn.cross_validation import train_test_split
+
+from pystruct.models import ChainCRF
+from pystruct.learners import SubgradientSSVM
+
+
+NETTALK = ("http://archive.ics.uci.edu/ml/machine-learning-databases/"
+           "undocumented/connectionist-bench/nettalk/nettalk.data")
+
+
+def features(word):
+    def end_features(curr_position):
+        return ("c[-1]={}".format(word[curr_position - 1]),
+                "c[+1]={}".format(word[curr_position]))
+
+    def middle_features(curr_position):
+        return ("c[-2]={}".format(word[curr_position - 2]),
+                "c[+2]={}".format(word[curr_position + 1]),
+                "c[-2:-1]={}".format(word[curr_position - 1:curr_position]),
+                "c[1:2]={}".format(word[curr_position + 1:curr_position + 3]))
+
+    return [end_features(pos) if pos == 1 or pos == len(word) - 1
+            else end_features(pos) + middle_features(pos)
+            for pos in xrange(1, len(word))]
+
+
+def nettalk_syl_to_split(syl):
+    syllables = [syl[k - 1] != '>' and syl[k] in ('>', '0', '1', '2')
+                 for k in xrange(1, len(syl))]
+    stress = [k == '1' for k in syl]
+    return syllables, stress
+
+
+def nettalk_line(line):
+    try:
+        word, phon, syl, cls = line.strip().split('\t')
+        syllable, stress = nettalk_syl_to_split(syl)
+    except ValueError:
+        word, syllable, stress = "", [], []
+    return features(word), syllable, stress
+
+
+def numbered_nb(y):
+    new_y = np.empty(len(y), dtype=np.int)
+    last_split = -1
+    for k, is_split in enumerate(y):
+        if is_split:
+            last_split = k
+        new_y[k] = k - last_split
+    return new_y
+
+
+if __name__ == '__main__':
+    url = urlopen(NETTALK)
+    for _ in xrange(10):  # skip header
+        url.readline()
+    lines = [nettalk_line(line) for line in url]
+    url.close()
+
+    X, y = zip(*((word, tag) for (word, tag, _) in lines if len(word)))
+    hasher = FeatureHasher(input_type='string', n_features=2**10,
+                           non_negative=True)
+    X = np.array([hasher.transform(instance) for instance in X])
+    y = np.array([numbered_nb(this_y) for this_y in y])
+
+    # The random state ensures that all labels are in the train and test set
+    X_train, X_test, y_train, y_test = train_test_split(X, y, train_size=0.2,
+                                                        random_state=1)
+
+    # train linear chain CRF
+    model = ChainCRF(inference_method=('ad3', dict(branch_and_bound=True)))
+    ssvm = SubgradientSSVM(model=model, verbose=1, C=100, max_iter=5)
+    ssvm.fit(X_train, y_train)
+    y_pred = ssvm.predict(X_test)
+    score = np.mean([np.all((y_t == 0) == (y_p == 0))
+                    for (y_t, y_p) in zip(y_test, y_pred)])
+    print("Test score: {:2.2f}".format(ssvm.score(X_test, y_test)))
diff --git a/pystruct/models/graph_crf.py b/pystruct/models/graph_crf.py
@@ -1,5 +1,7 @@
 import numpy as np
 
+from sklearn.utils.extmath import safe_sparse_dot
+
 from .crf import CRF
 from ..utils import expand_sym, compress_sym
 
@@ -117,7 +119,7 @@ def _get_unary_potentials(self, x, w):
         unary_params = w[:self.n_states * self.n_features].reshape(
             self.n_states, self.n_features)
 
-        return np.dot(features, unary_params.T)
+        return safe_sparse_dot(features, unary_params.T, dense_output=True)
 
     def psi(self, x, y):
         """Feature vector associated with instance (x, y).
@@ -165,7 +167,8 @@ def psi(self, x, y):
             pw = np.dot(unary_marginals[edges[:, 0]].T,
                         unary_marginals[edges[:, 1]])
 
-        unaries_acc = np.dot(unary_marginals.T, features)
+        unaries_acc = safe_sparse_dot(unary_marginals.T, features,
+                                      dense_output=True)
         if self.directed:
             pw = pw.ravel()
         else: