def test_lcwa_label_smoothing(self):
"""Test if output is correct for the LCWA training loop use case."""
# Create dummy dense labels
labels = torch.zeros(self.batch_size, self.num_entities)
for i in range(self.batch_size):
labels[i, self.random.randint(self.num_entities)] = 1.0
# Check if labels form a probability distribution
np.testing.assert_allclose(torch.sum(labels, dim=1).numpy(), 1.0)
# Apply label smoothing
smooth_labels = apply_label_smoothing(labels=labels,
epsilon=self.epsilon,
num_classes=self.num_entities)
# Check if smooth labels form probability distribution
np.testing.assert_allclose(torch.sum(smooth_labels, dim=1).numpy(),
1.0,
rtol=self.relative_tolerance)
def test_slcwa_label_smoothing(self):
"""Test if output is correct for the sLCWA training loop use case."""
# Create dummy sLCWA labels
ones = torch.ones(self.batch_size, 1)
zeros = torch.zeros(self.batch_size, 1)
labels = torch.cat([ones, zeros], dim=0)
# Apply label smoothing
smooth_labels = apply_label_smoothing(labels=labels,
epsilon=self.epsilon,
num_classes=self.num_entities)
exp_true = 1.0 - self.epsilon
np.testing.assert_allclose(smooth_labels[:self.batch_size],
exp_true,
rtol=self.relative_tolerance)
exp_false = self.epsilon / (self.num_entities - 1.)
np.testing.assert_allclose(smooth_labels[self.batch_size:],
exp_false,
rtol=self.relative_tolerance)