def main():
root = os.path.join(os.path.dirname(__file__), "../../")
dataset = MultiNLIDataset(root)
trainer = BaselineTrainer(root, log_dir="classifier_baseline")
trainer.build()
sequence_length = 25
vocab_size = len(trainer.preprocessor.vocabulary.get())
def preprocessor(x):
_x = trainer.preprocess(x, sequence_length)
return _x["text"]
model = MergeClassifier(vocab_size)
model.build(trainer.num_classes, preprocessor)
metrics = trainer.train(model.model,
epochs=25,
sequence_length=sequence_length,
representation="GloVe.6B.100d")
test_data = dataset.test_data()
y_pred = model.predict(test_data["text"])
print(
classification_report(test_data["label"],
y_pred,
target_names=dataset.labels()))
def main(graph_type="dependency", epochs=25):
root = os.path.join(os.path.dirname(__file__), "../../")
dataset = MultiNLIDataset(root)
if graph_type == "dependency":
graph_builder = DependencyGraph(lang="en")
elif graph_type == "similarity":
graph_builder = SimilarityGraph(lang="en")
else:
graph_builder = StaticGraph(lang="en")
trainer = Trainer(graph_builder, root, log_dir="classifier")
trainer.build()
sequence_length = 25
vocab_size = len(trainer.preprocessor.vocabulary.get())
def preprocessor(x):
_x = trainer.preprocess(x, sequence_length)
values = (_x["text"], _x["graph"])
return values
model = GraphBasedClassifier(vocab_size, sequence_length, lstm=None)
model.build(trainer.num_classes, preprocessor)
metrics = trainer.train(model.model, epochs=epochs)
test_data = dataset.test_data()
y_pred = model.predict(test_data["text"])
print(
classification_report(test_data["label"],
y_pred,
target_names=dataset.labels()))
def test_baseline(self):
root = os.path.join(os.path.dirname(__file__), "../../")
dataset = MultiNLIDataset(root)
data = dataset.test_data()
classifier = TfidfClassifier()
scores = classifier.fit(data["text"], data["label"])
self.assertTrue(len(scores) > 0)
def main():
root = os.path.join(os.path.dirname(__file__), "../../")
dataset = MultiNLIDataset(root)
classifier = TfidfClassifier()
train_data = dataset.train_data()
scores = classifier.fit(train_data["text"], train_data["label"])
test_data = dataset.test_data()
y_pred = classifier.predict(test_data["text"])
print(classification_report(test_data["label"], y_pred,
target_names=dataset.labels()))
def _test_train(self, graph_type):
root = os.path.join(os.path.dirname(__file__), "../../")
sequence_length = 25
heads = 3
dataset = MultiNLIDataset(root)
test_data = dataset.test_data()
index = np.random.randint(len(test_data), size=1)[0]
text = test_data["text"].iloc[index]
graph_builder = None
if graph_type == "dependency":
graph_builder = DependencyGraph(lang="en")
elif graph_type == "similarity":
graph_builder = SimilarityGraph(lang="en")
else:
graph_builder = StaticGraph(lang="en")
trainer = Trainer(graph_builder,
root,
preprocessor_name="test_ct_preprocessor")
trainer.build(data_kind="test")
def preprocessor(x):
_x = trainer.preprocess(x, sequence_length)
values = (_x["text"], _x["graph"])
return values
_, g = preprocessor([text])
vocab_size = len(trainer.preprocessor.vocabulary.get())
model = GraphBasedClassifier(vocab_size, sequence_length, heads=heads)
model.build(trainer.num_classes, preprocessor)
metrics = trainer.train(model.model, epochs=2)
os.remove(trainer.preprocessor_path)
self.assertTrue(
metrics.history["acc"][-1] - metrics.history["acc"][0] > 0)
attention = model.show_attention([text])
self.assertEqual(len(attention), 1) # batch size
attention = attention[0]
self.assertEqual(len(attention), 2) # layer count
attention = attention[0]
self.assertEqual(attention.shape,
(heads, sequence_length, sequence_length))
def test_download(self):
root = os.path.join(os.path.dirname(__file__), "../../")
dataset = MultiNLIDataset(root, prefix="test")
dataset.download()
train_data = dataset.train_data()
test_data = dataset.test_data()
for d in [train_data, test_data]:
self.assertTrue(len(d) > 0)
counts = d["label"].value_counts().values.tolist()
c = counts[0]
for _c in counts:
self.assertEqual(c, _c)
for k in ["train", "test"]:
self.assertTrue(os.path.exists(dataset.interim_file(k)))
os.remove(dataset.interim_file(k))
self.assertTrue(os.path.exists(dataset.processed_file(k)))
os.remove(dataset.processed_file(k))
def num_classes(self):
return len(MultiNLIDataset.labels())
def download(self):
r = MultiNLIDataset(self.storage.root).download()
return r