def test_is_binary():
t1 = torch.FloatTensor([0, 1, 0, 0])
t2 = torch.FloatTensor([0, -1, 0, 0])
t3 = torch.FloatTensor([0, 0.1, 0.2, 0])
assert is_binary(t1)
assert not is_binary(t2)
assert not is_binary(t3)
def __new__(cls, values, mask, left_justify=True):
if not isinstance(values, Variable) or not isinstance(mask, Variable):
raise ValueError('values and mask must both be of type Variable.')
m = mask.data
if len(m.size()) == 0:
raise ValueError('Mask must not be 0-dimensional')
# check that mask is binary
if not is_binary(m):
raise ValueError('Mask must be binary:\n{}'.format(mask))
# check that mask is left-justified
# since mask is binary, we just need to check that it is monotonically non-increasing from left to right
batch_size, seq_len = m.size()
if seq_len > 1 and left_justify:
diffs = m[:, 1:] - m[:, :-1] # (batch_size, max_seq_length - 1)
non_increasing = diffs <= 0
all_non_increasing = (torch.prod(non_increasing) == 1)
if not all_non_increasing:
raise ValueError(
'Mask must be left-justified:\n{}'.format(mask))
self = super(SequenceBatch, cls).__new__(cls, values, mask)
return self