def test_log_softmax_without_restrictions():
inputs = Variable(torch.randn(2, 5))
outputs = log_softmax(inputs)
assert isinstance(outputs, Variable)
assert outputs.size() == inputs.size()
assert all(x == pytest.approx(1.) for x in outputs.exp().sum(dim=1).data)
def test_log_softmax_with_restrictions():
restrictions = torch.LongTensor([0, 2])
inputs = Variable(torch.randn(1, 5))
outputs = log_softmax(inputs, restrictions=restrictions)
nonzero_indices = outputs.view(-1).exp().data.nonzero().view(-1)
assert nonzero_indices.tolist() == [1, 3, 4]
def test_log_softmax():
restrictions = torch.LongTensor([0, 2])
inputs = Variable(torch.randn(1, 5))
outputs = log_softmax(inputs, restrictions)
assert isinstance(outputs, Variable)
assert outputs.size() == (1, 5)
nonzero_indices = outputs.view(-1).exp().data.nonzero().view(-1)
assert all(nonzero_indices.eq(torch.LongTensor([1, 3, 4])))
def test_log_softmax_with_invalid_restrictions_dimension():
restrictions = torch.LongTensor([[0, 2]])
inputs = Variable(torch.randn(1, 5))
with pytest.raises(ValueError) as excinfo:
log_softmax(inputs, restrictions=restrictions)
assert 'restrictions must have dimension of 1, got 2' in str(excinfo.value)