def train_batch(model, xy):
x, y = xy
loss = 0.0
for x, y in zip(X, Y):
eg = Example(nr_class=model.nr_class, widths=model.widths)
eg.features = x
eg.costs = y
eg = list(model.update(eg, force_update=True))[-1]
loss += eg.loss
return loss
def get_scores(nr_class, model, feats):
eg = Example(nr_class)
eg.set_features(feats)
model(eg)
return list(eg.scores)
def get_score(nr_class, model, feats, clas):
eg = Example(nr_class)
eg.set_features(feats)
eg.set_label(clas)
model(eg)
return eg.scores[clas]
def test_basic():
nr_class = 3
model = AveragedPerceptron(((1,), (2,), (3,), (4,), (5,)))
instances = [
(1, {1: 1, 3: -5}),
(2, {2: 4, 3: 5})
]
for clas, feats in instances:
eg = Example(nr_class)
eg.set_features(feats)
model(eg)
eg.set_label(clas)
model.update(eg)
eg = Example(nr_class)
eg.set_features({1: 2, 2: 1})
model(eg)
assert eg.guess == 2
eg = Example(nr_class)
eg.set_features({0: 2, 2: 1})
model(eg)
assert eg.scores[1] == 0
eg = Example(nr_class)
eg.set_features({1: 2, 2: 1})
model(eg)
assert eg.scores[2] > 0
eg = Example(nr_class)
eg.set_features({1: 2, 1: 1})
model(eg)
assert eg.scores[1] > 0
eg = Example(nr_class)
eg.set_features({0: 3, 3: 1})
model(eg)
assert eg.scores[1] < 0
eg = Example(nr_class)
eg.set_features({0: 3, 3: 1})
model(eg)
assert eg.scores[2] > 0
def __init__(self, nlp, nr_class):
self.nlp = nlp
self.nr_class = nr_class
self._eg = Example(nr_class=nr_class)
self._model = AveragedPerceptron([])
def Eg(self, text, label=None):
bow = self.get_feats(text, dropout=bool(label is not None))
eg = Example(nr_class=self.nr_class, nr_feat=len(bow))
eg.costs = [i != label for i in range(self.nr_class)]
eg.features = bow
return eg
def test_basic():
nr_class = 3
model = AveragedPerceptron(((1, ), (2, ), (3, ), (4, ), (5, )))
instances = [(1, {1: 1, 3: -5}), (2, {2: 4, 3: 5})]
for clas, feats in instances:
eg = Example(nr_class)
eg.features = feats
model(eg)
eg.costs = [i != clas for i in range(nr_class)]
model.update(eg)
eg = Example(nr_class)
eg.features = {1: 2, 2: 1}
model(eg)
assert eg.guess == 2
eg = Example(nr_class)
eg.features = {0: 2, 2: 1}
model(eg)
assert eg.scores[1] == 0
eg = Example(nr_class)
eg.features = {1: 2, 2: 1}
model(eg)
assert eg.scores[2] > 0
eg = Example(nr_class)
eg.features = {1: 2, 1: 1}
model(eg)
assert eg.scores[1] > 0
eg = Example(nr_class)
eg.features = {0: 3, 3: 1}
model(eg)
assert eg.scores[1] < 0
eg = Example(nr_class)
eg.features = {0: 3, 3: 1}
model(eg)
assert eg.scores[2] > 0
def get_scores(nr_class, model, feats):
eg = Example(nr_class)
eg.features = feats
model(eg)
return list(eg.scores)
def get_score(nr_class, model, feats, clas):
eg = Example(nr_class)
eg.features = feats
eg.costs = [i != clas for i in range(nr_class)]
model(eg)
return eg.scores[clas]
def get_score(nr_class, model, feats, clas):
eg = Example(nr_class)
eg.features = feats
eg.costs = [i != clas for i in range(nr_class)]
model(eg)
return eg.scores[clas]
def test_basic():
nr_class = 3
model = AveragedPerceptron(((1,), (2,), (3,), (4,), (5,)))
instances = [
(1, {1: 1, 3: -5}),
(2, {2: 4, 3: 5})
]
for clas, feats in instances:
eg = Example(nr_class)
eg.features = feats
model(eg)
eg.costs = [i != clas for i in range(nr_class)]
model.update(eg)
eg = Example(nr_class)
eg.features = {1: 2, 2: 1}
model(eg)
assert eg.guess == 2
eg = Example(nr_class)
eg.features = {0: 2, 2: 1}
model(eg)
assert eg.scores[1] == 0
eg = Example(nr_class)
eg.features = {1: 2, 2: 1}
model(eg)
assert eg.scores[2] > 0
eg = Example(nr_class)
eg.features = {1: 2, 1: 1}
model(eg)
assert eg.scores[1] > 0
eg = Example(nr_class)
eg.features = {0: 3, 3: 1}
model(eg)
assert eg.scores[1] < 0
eg = Example(nr_class)
eg.features = {0: 3, 3: 1}
model(eg)
assert eg.scores[2] > 0
def test_embedding():
model = NeuralNet((10, 4, 2), embed=((5, ), (0, 0)), rho=0.0, eta=0.005)
assert model.nr_in == 10
eg = Example(nr_class=model.nr_class, widths=model.widths)
eg.features = {(0, 1): 2.5}
model(eg)
assert eg.activation(0, 0) != 0
assert eg.activation(0, 1) != 0
assert eg.activation(0, 2) != 0
assert eg.activation(0, 3) != 0
assert eg.activation(0, 4) != 0
eg = Example(nr_class=model.nr_class, widths=model.widths)
eg.features = {(1, 1867): 0.5}
model(eg)
assert eg.activation(0, 0) == 0.0
assert eg.activation(0, 1) == 0.0
assert eg.activation(0, 2) == 0.0
assert eg.activation(0, 3) == 0.0
assert eg.activation(0, 4) == 0.0
assert eg.activation(0, 5) != 0.0
assert eg.activation(0, 6) != 0.0
assert eg.activation(0, 7) != 0.0
assert eg.activation(0, 8) != 0.0
assert eg.activation(0, 9) != 0.0
