Fix x keras deep learning example

examples/keras_parikh_entailment/__main__.py

@@ -12,6 +12,11 @@ from spacy_hook import create_similarity_pipeline
 
 from keras_decomposable_attention import build_model
 
+try:
+    import cPickle as pickle
+except ImportError:
+    import pickle
+
 
 def train(model_dir, train_loc, dev_loc, shape, settings):
     train_texts1, train_texts2, train_labels = read_snli(train_loc)
@@ -19,6 +24,7 @@ def train(model_dir, train_loc, dev_loc, shape, settings):
 
     print("Loading spaCy")
     nlp = spacy.load('en')
+    assert nlp.path is not None
     print("Compiling network")
     model = build_model(get_embeddings(nlp.vocab), shape, settings)
     print("Processing texts...")
@@ -36,35 +42,41 @@ def train(model_dir, train_loc, dev_loc, shape, settings):
         validation_data=([dev_X1, dev_X2], dev_labels),
         nb_epoch=settings['nr_epoch'],
         batch_size=settings['batch_size'])
+    if not (nlp.path / 'similarity').exists():
+        (nlp.path / 'similarity').mkdir()
+    print("Saving to", model_dir / 'similarity')
+    weights = model.get_weights()
+    with (nlp.path / 'similarity' / 'model').open('wb') as file_:
+        pickle.dump(weights[1:], file_)
+    with (nlp.path / 'similarity' / 'config.json').open('wb') as file_:
+        file_.write(model.to_json())
 
 
 def evaluate(model_dir, dev_loc):
-    nlp = spacy.load('en', path=model_dir,
-            tagger=False, parser=False, entity=False, matcher=False,
+    dev_texts1, dev_texts2, dev_labels = read_snli(dev_loc)
+    nlp = spacy.load('en',
             create_pipeline=create_similarity_pipeline)
-    n = 0
-    correct = 0
-    for (text1, text2), label in zip(dev_texts, dev_labels):
+    total = 0.
+    correct = 0.
+    for text1, text2, label in zip(dev_texts1, dev_texts2, dev_labels):
         doc1 = nlp(text1)
         doc2 = nlp(text2)
         sim = doc1.similarity(doc2)
-        if bool(sim >= 0.5) == label:
+        if sim.argmax() == label.argmax():
             correct += 1
-        n += 1
+        total += 1
     return correct, total
 
 
 def demo(model_dir):
     nlp = spacy.load('en', path=model_dir,
-            tagger=False, parser=False, entity=False, matcher=False,
             create_pipeline=create_similarity_pipeline)
-    doc1 = nlp(u'Worst fries ever! Greasy and horrible...')
-    doc2 = nlp(u'The milkshakes are good. The fries are bad.')
-    print('doc1.similarity(doc2)', doc1.similarity(doc2))
-    sent1a, sent1b = doc1.sents
-    print('sent1a.similarity(sent1b)', sent1a.similarity(sent1b))
-    print('sent1a.similarity(doc2)', sent1a.similarity(doc2))
-    print('sent1b.similarity(doc2)', sent1b.similarity(doc2))
+    doc1 = nlp(u'What were the best crime fiction books in 2016?')
+    doc2 = nlp(
+        u'What should I read that was published last year? I like crime stories.')
+    print(doc1)
+    print(doc2)
+    print("Similarity", doc1.similarity(doc2))
 
 
 LABELS = {'entailment': 0, 'contradiction': 1, 'neutral': 2}
@@ -119,7 +131,8 @@ def main(mode, model_dir, train_loc, dev_loc,
     if mode == 'train':
         train(model_dir, train_loc, dev_loc, shape, settings)
     elif mode == 'evaluate':
-        evaluate(model_dir, dev_loc)
+        correct, total = evaluate(model_dir, dev_loc)
+        print(correct, '/', total, correct / total)
     else:
         demo(model_dir)
 

examples/keras_parikh_entailment/keras_decomposable_attention.py

@@ -12,6 +12,8 @@ from keras.models import Sequential, Model, model_from_json
 from keras.regularizers import l2
 from keras.optimizers import Adam
 from keras.layers.normalization import BatchNormalization
+from keras.layers.pooling import GlobalAveragePooling1D, GlobalMaxPooling1D
+from keras.layers import Merge
 
 
 def build_model(vectors, shape, settings):
@@ -176,8 +178,10 @@ class _Comparison(object):
 
     def __call__(self, sent, align, **kwargs):
         result = self.model(merge([sent, align], mode='concat')) # Shape: (i, n)
-        result = _GlobalSumPooling1D()(result, mask=self.words)
-        result = BatchNormalization()(result)
+        avged = GlobalAveragePooling1D()(result, mask=self.words)
+        maxed = GlobalMaxPooling1D()(result, mask=self.words)
+        merged = merge([avged, maxed])
+        result = BatchNormalization()(merged)
         return result
 
 
@@ -261,6 +265,4 @@ def test_fit_model():
               batch_size=4)
 
 
-
-
 __all__ = [build_model]

examples/keras_parikh_entailment/spacy_hook.py

@@ -1,33 +1,40 @@
 from keras.models import model_from_json
 import numpy
 import numpy.random
+import json
+from spacy.tokens.span import Span
+
+try:
+    import cPickle as pickle
+except ImportError:
+    import pickle
 
 
 class KerasSimilarityShim(object):
     @classmethod
-    def load(cls, path, nlp, get_features=None):
+    def load(cls, path, nlp, get_features=None, max_length=100):
         if get_features is None:
-            get_features = doc2ids
+            get_features = get_word_ids
         with (path / 'config.json').open() as file_:
-            config = json.load(file_)
-        model = model_from_json(config['model'])
+            model = model_from_json(file_.read())
         with (path / 'model').open('rb') as file_:
             weights = pickle.load(file_)
         embeddings = get_embeddings(nlp.vocab)
         model.set_weights([embeddings] + weights)
-        return cls(model, get_features=get_features)
+        return cls(model, get_features=get_features, max_length=max_length)
 
-    def __init__(self, model, get_features=None):
+    def __init__(self, model, get_features=None, max_length=100):
         self.model = model
         self.get_features = get_features
+        self.max_length = max_length
 
     def __call__(self, doc):
         doc.user_hooks['similarity'] = self.predict
         doc.user_span_hooks['similarity'] = self.predict
 
     def predict(self, doc1, doc2):
-        x1 = self.get_features(doc1)
-        x2 = self.get_features(doc2)
+        x1 = self.get_features([doc1], max_length=self.max_length, tree_truncate=True)
+        x2 = self.get_features([doc2], max_length=self.max_length, tree_truncate=True)
         scores = self.model.predict([x1, x2])
         return scores[0]
 
@@ -45,7 +52,10 @@ def get_word_ids(docs, rnn_encode=False, tree_truncate=False, max_length=100, nr
     Xs = numpy.zeros((len(docs), max_length), dtype='int32')
     for i, doc in enumerate(docs):
         if tree_truncate:
-            queue = [sent.root for sent in doc.sents]
+            if isinstance(doc, Span):
+                queue = [doc.root]
+            else:
+                queue = [sent.root for sent in doc.sents]
         else:
             queue = list(doc)
         words = []
@@ -71,7 +81,9 @@ def get_word_ids(docs, rnn_encode=False, tree_truncate=False, max_length=100, nr
 
 
 def create_similarity_pipeline(nlp):
-    return [SimilarityModel.load(
-                nlp.path / 'similarity',
-                nlp,
-                feature_extracter=get_features)]
+    return [
+        nlp.tagger,
+        nlp.entity,
+        nlp.parser,
+        KerasSimilarityShim.load(nlp.path / 'similarity', nlp, max_length=10)
+    ]