def test_inference(self):
# perfect annotation, check that inferred label is correct
nclasses, nannotators, nitems = 3, 5, 50*8
# create random model (this is our ground truth model)
gamma = np.ones((nclasses,)) / float(nclasses)
theta = np.ones((8,)) * 0.999
true_model = ModelBt(nclasses, nannotators, gamma, theta)
# create random data
labels = true_model.generate_labels(nitems)
annotations = true_model.generate_annotations_from_labels(labels)
posterior = true_model.infer_labels(annotations)
testing.assert_allclose(posterior.sum(1), 1., atol=1e-6, rtol=0.)
inferred = posterior.argmax(1)
testing.assert_equal(inferred, labels)
self.assertTrue(np.all(posterior[np.arange(nitems),inferred] > 0.999))
# at chance annotation, disagreeing annotators: get back prior
gamma = ModelBt._random_gamma(nclasses)
theta = np.ones((nannotators,)) / float(nclasses)
model = ModelBt(nclasses, nannotators, gamma, theta)
data = np.array([[MV, 0, 1, 2, MV]])
testing.assert_almost_equal(np.squeeze(model.infer_labels(data)),
model.gamma, 6)
def test_inference(self):
# perfect annotation, check that inferred label is correct
nclasses, nannotators, nitems = 3, 5, 50 * 8
# create random model (this is our ground truth model)
gamma = np.ones((nclasses, )) / float(nclasses)
theta = np.ones((8, )) * 0.999
true_model = ModelBt(nclasses, nannotators, gamma, theta)
# create random data
labels = true_model.generate_labels(nitems)
annotations = true_model.generate_annotations_from_labels(labels)
posterior = true_model.infer_labels(annotations)
testing.assert_allclose(posterior.sum(1), 1., atol=1e-6, rtol=0.)
inferred = posterior.argmax(1)
testing.assert_equal(inferred, labels)
self.assertTrue(np.all(posterior[np.arange(nitems), inferred] > 0.999))
# at chance annotation, disagreeing annotators: get back prior
gamma = ModelBt._random_gamma(nclasses)
theta = np.ones((nannotators, )) / float(nclasses)
model = ModelBt(nclasses, nannotators, gamma, theta)
data = np.array([[MV, 0, 1, 2, MV]])
testing.assert_almost_equal(np.squeeze(model.infer_labels(data)),
model.gamma, 6)
