Skip to content

fix: normalize float16 embeddings without NaNs #368

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
xblwh wants to merge 1 commit into from
+21 −3
Closed

fix: normalize float16 embeddings without NaNs #368

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
3 changes: 2 additions & 1 deletion model2vec/model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -476,7 +476,8 @@ def _encode_batch(self, sentences: Sequence[str], normalize: bool) -> np.ndarray
out[i] = emb.mean(axis=0)

if normalize:
norm = np.linalg.norm(out, axis=1, keepdims=True) + 1e-32
norm_input = out.astype(np.float32) if out.dtype == np.float16 else out
norm = np.linalg.norm(norm_input, axis=1, keepdims=True) + 1e-32
np.divide(out, norm, out=out)

return out
Expand Down
21 changes: 19 additions & 2 deletions tests/test_model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -42,11 +42,12 @@ def test_encode_single_sentence(
assert encoded.shape == (2,)


@pytest.mark.parametrize("dtype", ["float16", "float32", "float64"])
def test_encode_single_sentence_empty(
mock_vectors: np.ndarray, mock_tokenizer: Tokenizer, mock_config: dict[str, str]
mock_vectors: np.ndarray, mock_tokenizer: Tokenizer, mock_config: dict[str, str], dtype: str
) -> None:
"""Test encoding of a single empty sentence."""
model = StaticModel(vectors=mock_vectors, tokenizer=mock_tokenizer, config=mock_config)
model = StaticModel(vectors=mock_vectors.astype(dtype), tokenizer=mock_tokenizer, config=mock_config)
model.normalize = True
encoded = model.encode("")
assert not np.isnan(encoded).any()
Expand Down Expand Up @@ -159,6 +160,22 @@ def test_normalize(mock_vectors: np.ndarray, mock_tokenizer: Tokenizer, mock_con
np.testing.assert_almost_equal(normalized, expected)


@pytest.mark.parametrize("dtype", ["float16", "float32", "float64"])
@pytest.mark.parametrize("scale", [1e-4, 1.0, 1e4])
def test_normalize_zero_and_nonzero_embeddings(
mock_vectors: np.ndarray, mock_tokenizer: Tokenizer, dtype: str, scale: float
) -> None:
"""Keep zero vectors finite and normalize nonzero vectors across precisions and scales."""
model = StaticModel(vectors=(mock_vectors * scale).astype(dtype), tokenizer=mock_tokenizer, normalize=True)
with np.errstate(divide="raise", invalid="raise", over="raise"):
encoded = model.encode(["", "unknown", "word3", "word1 word2"])

assert encoded.dtype == np.dtype(dtype)
assert np.isfinite(encoded).all()
np.testing.assert_array_equal(encoded[:3], np.zeros((3, model.dim)))
assert np.linalg.norm(encoded[3].astype(np.float64)) == pytest.approx(1.0, abs=1e-3)


def test_save_pretrained(
tmp_path: Path, mock_vectors: np.ndarray, mock_tokenizer: Tokenizer, mock_config: dict[str, str]
) -> None:
Expand Down