def test_sampled_softmax_has_greater_loss_in_train_mode(self):
sampled_softmax = SampledSoftmaxLoss(num_words=10000, embedding_dim=12, num_samples=10)
# sequence_length, embedding_dim
embedding = torch.rand(100, 12)
targets = torch.randint(0, 1000, (100,)).long()
sampled_softmax.train()
train_loss = sampled_softmax(embedding, targets).item()
sampled_softmax.eval()
eval_loss = sampled_softmax(embedding, targets).item()
assert eval_loss > train_loss
def test_sampled_equals_unsampled_when_biased_against_non_sampled_positions(
self):
sampled_softmax = SampledSoftmaxLoss(num_words=10000,
embedding_dim=12,
num_samples=10)
unsampled_softmax = SoftmaxLoss(num_words=10000, embedding_dim=12)
# fake out choice function
FAKE_SAMPLES = [100, 200, 300, 400, 500, 600, 700, 800, 900, 9999]
def fake_choice(num_words: int,
num_samples: int) -> Tuple[np.ndarray, int]:
assert (num_words, num_samples) == (10000, 10)
return np.array(FAKE_SAMPLES), 12
sampled_softmax.choice_func = fake_choice
# bias out the unsampled terms:
for i in range(10000):
if i not in FAKE_SAMPLES:
unsampled_softmax.softmax_b[i] = -10000
# set weights equal, use transpose because opposite shapes
sampled_softmax.softmax_w.data = unsampled_softmax.softmax_w.t()
sampled_softmax.softmax_b.data = unsampled_softmax.softmax_b
sampled_softmax.train()
unsampled_softmax.train()
# sequence_length, embedding_dim
embedding = torch.rand(100, 12)
targets = torch.randint(0, 1000, (100, )).long()
full_loss = unsampled_softmax(embedding, targets).item()
sampled_loss = sampled_softmax(embedding, targets).item()
# Should be close
pct_error = (sampled_loss - full_loss) / full_loss
assert abs(pct_error) < 0.001
