def probability_ratio(
self, tag: str, x: ep.Tensor, y: ep.Tensor, step: int
) -> None:
x_ = x.float32().mean(axis=0).item()
y_ = y.float32().mean(axis=0).item()
if y_ == 0:
return
self.writer.add_scalar(tag, x_ / y_, step)
def test_logical_and_nonboolean(t: Tensor, f: Callable[[Tensor, Tensor],
Tensor]) -> None:
t = t.float32()
f(t > 1, t > 1)
with pytest.raises(ValueError):
f(t, t > 1)
with pytest.raises(ValueError):
f(t > 1, t)
with pytest.raises(ValueError):
f(t, t)
def apply_decision_rule(
decision_rule: str,
beta: float,
best_advs: ep.Tensor,
best_advs_norms: ep.Tensor,
x_k: ep.Tensor,
x_0: ep.Tensor,
found_advs: ep.Tensor,
):
if decision_rule == "EN":
norms = beta * flatten(x_k - x_0).abs().sum(
axis=-1) + flatten(x_k - x_0).square().sum(axis=-1)
elif decision_rule == "L1":
norms = flatten(x_k - x_0).abs().sum(axis=-1)
else:
raise ValueError("invalid decision rule")
new_best = (norms < best_advs_norms).float32() * found_advs.float32()
new_best = atleast_kd(new_best, best_advs.ndim)
best_advs = new_best * x_k + (1 - new_best) * best_advs
best_advs_norms = ep.minimum(norms, best_advs_norms)
return best_advs, best_advs_norms
def probability(self, tag: str, x: ep.Tensor, step: int) -> None:
self.writer.add_scalar(tag, x.float32().mean(axis=0).item(), step)