def _predict_transition_probabilities_numpy(
X: np.ndarray,
W: np.ndarray,
softmax_scale: float = 1.0,
center_mean: bool = True,
scale_by_norm: bool = True,
) -> Tuple[np.ndarray, np.ndarray]:
if center_mean:
# pearson correlation
W -= np.expand_dims(np_mean(W, axis=1), axis=1)
if X.shape[0] == 1:
if center_mean:
# pearson correlation
X = X - np.mean(X)
if scale_by_norm:
# cosine or pearson correlation
denom = np.linalg.norm(X) * norm(W, axis=1)
mask = denom == 0
denom[mask] = 1
return _softmax_masked(W.dot(X[0]) / denom, mask, softmax_scale)
return _softmax(W.dot(X[0]), softmax_scale)
assert X.shape[0] == W.shape[0], "Wrong shape."
if center_mean:
X = X - np.expand_dims(np_mean(X, axis=1), axis=1)
if scale_by_norm:
denom = norm(X, axis=1) * norm(W, axis=1)
mask = denom == 0
denom[mask] = 1
return _softmax_masked(
np.array([np.dot(X[i], W[i]) for i in range(X.shape[0])]) / denom,
mask,
softmax_scale,
)
return _softmax(np.array([np.dot(X[i], W[i]) for i in range(X.shape[0])]),
softmax_scale)
def test_numba_norm_axis_1(self, seed):
np.random.seed(seed)
x = np.random.normal(size=(10, 10))
np.testing.assert_allclose(np.linalg.norm(x, axis=1), norm(x, 1))