def test_feasible(Constraint, alpha):
"""Tests if prox and LMO yield feasible points"""
# TODO: implement feasibility check method in each constraint.
if Constraint == chop.constraints.GroupL1Ball:
groups = group_patches(x_patch_size=2, y_patch_size=2, x_image_size=6, y_image_size=6)
constraint = Constraint(alpha, groups)
elif Constraint == chop.constraints.Cone:
directions = torch.rand(2, 3, 6, 6)
cos_alpha = .2
constraint = Constraint(directions, cos_alpha)
elif Constraint == chop.constraints.Box:
constraint = Constraint(-1., 10.)
else:
constraint = Constraint(alpha)
for _ in range(10):
try:
data = (alpha + 1) * torch.rand(2, 3, 6, 6)
assert constraint.is_feasible(constraint.prox(data)).all()
except AttributeError: # Constraint doesn't have a prox operator
pass
try:
grad = (alpha + 1) * torch.rand(2, 3, 6, 6)
update_dir, _ = constraint.lmo(-grad, data)
s = update_dir + data
assert constraint.is_feasible(s).all()
except AttributeError: # Constraint doesn't have an LMO
pass
def test_groupL1Prox():
batch_size = 2
alpha = 10
groups = group_patches(x_patch_size=2, y_patch_size=2, x_image_size=6, y_image_size=6)
constraint = chop.constraints.GroupL1Ball(alpha, groups)
data = torch.rand(batch_size, 3, 6, 6)
constraint.prox(-data, step_size=.3)
def test_GroupL1LMO():
batch_size = 2
alpha = 1.
groups = group_patches(x_patch_size=2, y_patch_size=2, x_image_size=6, y_image_size=6)
constraint = chop.constraints.GroupL1Ball(alpha, groups)
data = torch.rand(batch_size, 3, 6, 6)
grad = torch.rand(batch_size, 3, 6, 6)
constraint.lmo(-grad, data)
def test_projections(constraint, alpha):
"""Tests that projections are true projections:
..math::
p\circ p = p
"""
batch_size = 8
if constraint == chop.constraints.GroupL1Ball:
groups = group_patches()
prox = constraint(alpha, groups).prox
elif constraint == chop.constraints.Cone:
directions = torch.rand(batch_size, 3, 32, 32)
prox = constraint(directions, cos_angle=.2).prox
else:
prox = constraint(alpha).prox
for _ in range(10):
data = torch.rand(batch_size, 3, 32, 32)
proj_data = prox(data)
# SVD reconstruction doesn't do better than 1e-5
double_proj = prox(proj_data)
assert double_proj.allclose(proj_data, atol=1e-5), (double_proj, proj_data)
normalize=True,
title=f'L{constraint.p}',
negative=False)
print(
f"F1 score: {f1_score(target.detach().cpu(), adv_labels.detach().cpu(), average='macro'):.3f}"
f" for alpha={alpha:.4f}")
print("GroupL1 constraint.")
# CIFAR-10
# groups = group_patches(x_patch_size=8, y_patch_size=8, x_image_size=32, y_image_size=32)
# Imagenet
groups = group_patches(x_patch_size=28,
y_patch_size=28,
x_image_size=224,
y_image_size=224)
for eps in [5e-2]:
alpha = eps * len(groups)
constraint_group = chop.constraints.GroupL1Ball(alpha, groups)
adversary_group = chop.Adversary(chop.optim.minimize_frank_wolfe)
# callback_group = Trace(callable=lambda kw: criterion(model(data + kw['x']), target))
callback_group = Trace()
_, delta_group = adversary_group.perturb(data,
target,
model,
criterion,
lmo=constraint_group.lmo,
