def test_ignore_row(self):
# If a sample has no valid targets, it should be ignored in the reduction.
config = self._get_config()
crit = SoftTargetCrossEntropyLoss.from_config(config)
outputs = torch.tensor([[1.0, 7.0, 0.0, 0.0, 2.0],
[4.0, 2.0, 1.0, 6.0, 0.5]])
targets = torch.tensor([[1, 0, 0, 0, 1], [-1, -1, -1, -1, -1]])
self.assertAlmostEqual(crit(outputs, targets).item(), self._get_loss())
def test_soft_target_cross_entropy(self):
config = self._get_config()
crit = SoftTargetCrossEntropyLoss.from_config(config)
outputs = self._get_outputs()
targets = self._get_targets()
self.assertAlmostEqual(crit(outputs, targets).item(), self._get_loss())
# Verify ignore index works
outputs = self._get_outputs()
targets = torch.tensor([[-1, 0, 0, 0, 1]])
self.assertAlmostEqual(crit(outputs, targets).item(), 5.01097918)
def test_soft_target_cross_entropy_none_reduction(self):
# reduction mode is "none"
config = self._get_config()
config["reduction"] = "none"
crit = SoftTargetCrossEntropyLoss.from_config(config)
outputs = torch.tensor([[1.0, 7.0, 0.0, 0.0, 2.0],
[4.0, 2.0, 1.0, 6.0, 0.5]])
targets = torch.tensor([[1, 0, 0, 0, 1], [0, 1, 0, 1, 0]])
loss = crit(outputs, targets)
self.assertEqual(loss.numel(), outputs.size(0))
def test_unnormalized_soft_target_cross_entropy(self):
config = {
"name": "soft_target_cross_entropy",
"ignore_index": -1,
"reduction": "mean",
"normalize_targets": False,
}
crit = SoftTargetCrossEntropyLoss.from_config(config)
outputs = self._get_outputs()
targets = self._get_targets()
self.assertAlmostEqual(crit(outputs, targets).item(), 11.0219593)
# Verify ignore index works
outputs = self._get_outputs()
targets = torch.tensor([[-1, 0, 0, 0, 1]])
self.assertAlmostEqual(crit(outputs, targets).item(), 5.01097965)
def test_soft_target_cross_entropy_integer_label(self):
config = self._get_config()
crit = SoftTargetCrossEntropyLoss.from_config(config)
outputs = self._get_outputs()
targets = torch.tensor([4])
self.assertAlmostEqual(crit(outputs, targets).item(), 5.01097918)