def test_loss():
d_scores = CategoricalCrossentropy().get_grad(scores0, labels0)
assert d_scores.dtype == "float32"
assert d_scores.shape == scores0.shape
d_scores = SequenceCategoricalCrossentropy().get_grad([scores0], [labels0])
assert d_scores[0].dtype == "float32"
assert d_scores[0].shape == scores0.shape
assert SequenceCategoricalCrossentropy().get_grad([], []) == []
def test_categorical_crossentropy(guesses, labels):
d_scores = CategoricalCrossentropy(normalize=True).get_grad(guesses, labels)
assert d_scores.shape == guesses.shape
# The normalization divides the difference (e.g. 0.4) by the number of vectors (4)
assert d_scores[1][0] == pytest.approx(0.1, eps)
assert d_scores[1][1] == pytest.approx(-0.1, eps)
# The third vector predicted all labels, but only the first one was correct
assert d_scores[2][0] == pytest.approx(0, eps)
assert d_scores[2][1] == pytest.approx(0.25, eps)
assert d_scores[2][2] == pytest.approx(0.25, eps)
# The fourth vector predicted no labels but should have predicted the last one
assert d_scores[3][0] == pytest.approx(0, eps)
assert d_scores[3][1] == pytest.approx(0, eps)
assert d_scores[3][2] == pytest.approx(-0.25, eps)
loss = CategoricalCrossentropy(normalize=True).get_loss(guesses, labels)
assert loss == pytest.approx(0.239375, eps)
def test_crossentropy_incorrect_scores_targets():
labels = numpy.asarray([2])
guesses_neg = numpy.asarray([[-0.1, 0.5, 0.6]])
with pytest.raises(ValueError, match=r"Cannot calculate.*guesses"):
CategoricalCrossentropy(normalize=True).get_grad(guesses_neg, labels)
guesses_larger_than_one = numpy.asarray([[1.1, 0.5, 0.6]])
with pytest.raises(ValueError, match=r"Cannot calculate.*guesses"):
CategoricalCrossentropy(normalize=True).get_grad(
guesses_larger_than_one, labels)
guesses_ok = numpy.asarray([[0.1, 0.4, 0.5]])
targets_neg = numpy.asarray([[-0.1, 0.5, 0.6]])
with pytest.raises(ValueError, match=r"Cannot calculate.*truth"):
CategoricalCrossentropy(normalize=True).get_grad(
guesses_ok, targets_neg)
targets_larger_than_one = numpy.asarray([[2.0, 0.5, 0.6]])
with pytest.raises(ValueError, match=r"Cannot calculate.*truth"):
CategoricalCrossentropy(normalize=True).get_grad(
guesses_ok, targets_larger_than_one)
def test_categorical_crossentropy_int_list_missing(guesses, labels):
d_scores = CategoricalCrossentropy(normalize=True,
missing_value=0).get_grad(
guesses, labels)
assert d_scores.shape == guesses.shape
# The normalization divides the difference (e.g. 0.4) by the number of vectors (4)
assert d_scores[1][0] == pytest.approx(0.1, eps)
assert d_scores[1][1] == pytest.approx(-0.1, eps)
# Label 0 is masked, because it represents the missing value
assert d_scores[2][0] == 0.0
assert d_scores[2][1] == 0.0
assert d_scores[2][2] == 0.0
# The fourth vector predicted no labels but should have predicted the last one
assert d_scores[3][0] == pytest.approx(0, eps)
assert d_scores[3][1] == pytest.approx(0, eps)
assert d_scores[3][2] == pytest.approx(-0.25, eps)
loss = CategoricalCrossentropy(normalize=True,
missing_value=0).get_loss(guesses, labels)
assert loss == pytest.approx(0.114375, eps)
def test_categorical_crossentropy_missing(guesses, labels):
d_scores = CategoricalCrossentropy(normalize=True).get_grad(
guesses, labels)
assert d_scores.shape == guesses.shape
eps = 0.0001
def test_loss():
d_scores = CategoricalCrossentropy().get_grad(scores0, labels0)
assert d_scores.dtype == "float32"
assert d_scores.shape == scores0.shape
d_scores = SequenceCategoricalCrossentropy().get_grad([scores0], [labels0])
assert d_scores[0].dtype == "float32"
assert d_scores[0].shape == scores0.shape
assert SequenceCategoricalCrossentropy().get_grad([], []) == []
@pytest.mark.parametrize("dist", [
CategoricalCrossentropy(),
CosineDistance(ignore_zeros=True),
L2Distance()
])
@pytest.mark.parametrize("vect", [scores0, guesses1, guesses2])
def test_equality(dist, vect):
assert int(dist.get_grad(vect, vect)[0][0]) == pytest.approx(0, eps)
assert dist.get_loss(vect, vect) == pytest.approx(0, eps)
@pytest.mark.parametrize("guesses, labels", [(guesses1, labels1),
(guesses1, labels1_full)])
def test_categorical_crossentropy(guesses, labels):
d_scores = CategoricalCrossentropy(normalize=True).get_grad(
guesses, labels)
assert d_scores.shape == guesses.shape
eps = 0.0001
def test_loss():
d_scores = CategoricalCrossentropy().get_grad(scores0, labels0)
assert d_scores.dtype == "float32"
assert d_scores.shape == scores0.shape
d_scores = SequenceCategoricalCrossentropy().get_grad([scores0], [labels0])
assert d_scores[0].dtype == "float32"
assert d_scores[0].shape == scores0.shape
assert SequenceCategoricalCrossentropy().get_grad([], []) == []
@pytest.mark.parametrize(
"dist", [CategoricalCrossentropy(), CosineDistance(ignore_zeros=True), L2Distance()]
)
@pytest.mark.parametrize("vect", [scores0, guesses1, guesses2])
def test_equality(dist, vect):
assert int(dist.get_grad(vect, vect)[0][0]) == pytest.approx(0, eps)
assert dist.get_loss(vect, vect) == pytest.approx(0, eps)
@pytest.mark.parametrize(
"guesses, labels", [(guesses1, labels1), (guesses1, labels1_full)]
)
def test_categorical_crossentropy(guesses, labels):
d_scores = CategoricalCrossentropy(normalize=True).get_grad(guesses, labels)
assert d_scores.shape == guesses.shape
# The normalization divides the difference (e.g. 0.4) by the number of vectors (4)