def test_neg_log_loss_batch_stable(generate_examples_and_batch): i1, t1, l1, i2, t2, l2, i, t, l = generate_examples_and_batch h = i1.size(2) crf = CRF(h, batch_first=False) crf.transitions_p.data = torch.rand(1, h, h) nll1 = crf.neg_log_loss(i1, t1, l1) nll2 = crf.neg_log_loss(i2, t2, l2) one_x_one = (nll1 + nll2) / 2 batched = crf.neg_log_loss(i, t, l) np.testing.assert_allclose(one_x_one.detach().numpy(), batched.detach().numpy())
def test_neg_log_loss(generate_batch):
unary, tags, lengths = generate_batch
h = unary.size(2)
crf = CRF(h, batch_first=False)
trans = torch.rand(h, h)
crf.transitions_p.data = trans.unsqueeze(0)
nll = crf.neg_log_loss(unary, tags, lengths)
new_trans = build_trans(trans)
unary = unary.transpose(0, 1)
tags = tags.transpose(0, 1)
scores = []
for u, t, l in zip(unary, tags, lengths):
emiss = build_emission(u[:l])
golds = t[:l].tolist()
scores.append(explicit_nll(emiss, new_trans, golds, Offsets.GO, Offsets.EOS))
gold_scores = np.mean(np.array(scores))
np.testing.assert_allclose(nll.detach().numpy(), gold_scores, rtol=1e-6)
def test_mask_same_after_update(generate_batch):
from torch.optim import SGD
unary, tags, lengths = generate_batch
h = unary.size(2)
constraint = torch.rand(h, h) < 0.5
crf = CRF(h, constraint_mask=constraint, batch_first=False)
opt = SGD(crf.parameters(), lr=10)
m1 = crf.constraint_mask.numpy()
t1 = crf.transitions_p.detach().clone().numpy()
l = crf.neg_log_loss(unary, tags, lengths)
l = torch.mean(l)
l.backward()
opt.step()
m2 = crf.constraint_mask.numpy()
t2 = crf.transitions_p.detach().numpy()
np.testing.assert_allclose(m1, m2)
with pytest.raises(AssertionError):
np.testing.assert_allclose(t1, t2)
