def test_two_means_invalid_dims(self):
two_means = TwoMeansBiasDirection()
with pytest.raises(ConfigurationError):
two_means(torch.zeros(2), torch.zeros(2))
with pytest.raises(ConfigurationError):
two_means(torch.zeros(2, 2), torch.zeros(2, 3))
def setup_method(self):
super().setup_method()
# embedding data from VERB demo
emb_filename = str(self.FIXTURES_ROOT / "fairness" /
"bias_embeddings.json")
with open(emb_filename) as emb_file:
emb_data = json.load(emb_file)
seed_embeddings1 = torch.cat([
torch.Tensor(emb_data["he"]).reshape(1, -1),
torch.Tensor(emb_data["him"]).reshape(1, -1),
])
seed_embeddings2 = torch.cat([
torch.Tensor(emb_data["she"]).reshape(1, -1),
torch.Tensor(emb_data["her"]).reshape(1, -1),
])
tm = TwoMeansBiasDirection()
self.bias_direction = tm(seed_embeddings1, seed_embeddings2)
evaluation_embeddings = []
expected_bias_mitigated_embeddings = []
for word in ["engineer", "banker", "nurse", "receptionist"]:
evaluation_embeddings.append(
torch.Tensor(emb_data[word]).reshape(1, -1))
expected_bias_mitigated_embeddings.append(
torch.Tensor(emb_data["linear_two_means_" + word]).reshape(
1, -1))
self.evaluation_embeddings = torch.cat(evaluation_embeddings).reshape(
2, 2, -1)
self.expected_bias_mitigated_embeddings = torch.cat(
expected_bias_mitigated_embeddings).reshape(2, 2, -1)
def test_two_means_with_grad(self, device: str):
seed_embeddings1 = torch.eye(2, device=device)
seed_embeddings2 = 1 - torch.eye(2, device=device)
seed_embeddings1 = seed_embeddings1.requires_grad_()
seed_embeddings2 = seed_embeddings2.requires_grad_()
assert seed_embeddings1.grad is None
assert seed_embeddings2.grad is None
two_means = TwoMeansBiasDirection(requires_grad=True)
test_bias_direction = two_means(seed_embeddings1, seed_embeddings2)
test_bias_direction.sum().backward()
assert seed_embeddings1.grad is not None
assert seed_embeddings2.grad is not None
def test_two_means_without_grad(self, device: str):
seed_embeddings1 = torch.eye(2, device=device)
seed_embeddings2 = 1 - torch.eye(2, device=device)
two_means = TwoMeansBiasDirection()
expected_bias_direction = torch.tensor(
[float("nan"), float("nan")], device=device)
test_bias_direction = two_means(seed_embeddings1, seed_embeddings2)
assert allclose(expected_bias_direction,
test_bias_direction,
equal_nan=True)
assert seed_embeddings1.grad is None
assert seed_embeddings2.grad is None
def __init__(
self,
seed_word_pairs_file: Union[PathLike, str],
tokenizer: Tokenizer,
direction_vocab: Optional[Vocabulary] = None,
namespace: str = "tokens",
requires_grad: bool = False,
noise: float = 1e-10,
):
self.ids1, self.ids2 = load_word_pairs(seed_word_pairs_file, tokenizer,
direction_vocab, namespace)
self.direction = TwoMeansBiasDirection(requires_grad=requires_grad)
self.noise = noise