def test_accuracy():
accuracy = Accuracy()
accuracy_with_logits = AccuracyWithLogits()
top2 = TopKAccuracy(2)
top3 = TopKAccuracy(3)
fake_accuracy = AverageMetric(accuracy.score)
# First try with a multi class input
x = J.Variable(J.Tensor([[0.9, 0, 0.1, 0], [0.2, 0.3, 0.1, 0.4]]))
# Labels are the indicies
y = J.Variable(J.LongTensor([0, 0]))
assert accuracy(x, y).item() == 50.
assert fake_accuracy(x, y).item() == 50.
# Since applying the softmax again won't change the ordering
assert accuracy_with_logits(x, y) == 50.
assert top2(x, y) == 50.
assert top3(x, y) == 100.
# Now try with binary class input
x_logit = J.Variable(J.Tensor([[100.], [-100.]]))
x = torch.sigmoid(x_logit)
y = J.Variable(J.LongTensor([0, 0]))
assert accuracy(x, y) == 50.
assert fake_accuracy(x, y) == 50.
assert accuracy_with_logits(x_logit, y) == 50.
def test_accumulation():
accuracy = Accuracy()
x = J.Variable(
J.Tensor([[0.9, 0, 0.1, 0], [0.2, 0.3, 0.1, 0.4], [1.0, 0.0, 0.0,
0.0]]))
y = J.Variable(J.LongTensor([1, 2, 3]))
accuracy(x, y)
x = J.Variable(J.Tensor([[0.9, 0, 0.1, 0], [0.2, 0.3, 0.1, 0.4]]))
y = J.Variable(J.LongTensor([0, 0]))
accuracy(x, y)
assert accuracy.accumulate() == 20.
accuracy = accuracy.reset()
accuracy(x, y)
assert accuracy.accumulate() == 50.
def multi_bce_with_logits_seq(outputs, targets, size_average=True):
flat_outputs = torch.cat(outputs, dim=0).unsqueeze(1) # B*Li x 1
flat_targets = Variable(J.Tensor(np.zeros(
(flat_outputs.size(0), )))) # B*Li
start = 0
for i, output in enumerate(outputs):
end = start + output.size(0)
flat_targets[start:end] = targets[i].expand(output.size(0))
start = end
return bce_with_logits(flat_outputs,
flat_targets,
size_average=size_average)
def test_categorical_crossentropy():
x = J.Variable(J.Tensor([[0.5, 0, 0.5, 0]]))
y = J.Variable(J.LongTensor([0]))
np.testing.assert_almost_equal(0.6931471824645996,
categorical_crossentropy(x, y).item())
def test_flatten_3d_tensor():
test_array = np.arange(2 * 3 * 4).reshape((2, 3, 4)).astype(np.float32)
expected_array = np.arange(2 * 3 * 4).reshape(2, 3 * 4).astype(np.float32)
test_tensor = J.Tensor(test_array)
actual_array = J.flatten(test_tensor).cpu().numpy()
np.testing.assert_array_equal(actual_array, expected_array)