changing attack running parameter and defense error type

advertorch/utils.py

@@ -222,6 +222,8 @@ def batch_l1_proj_flat(x, z=1):
 
     # Getting the elements to project in the batch
     indexes_b = torch.nonzero(v > z).view(-1)
+    if isinstance(z, torch.Tensor):
+        z[indexes_b][:, None]
     x_b = x[indexes_b]
     batch_size_b = x_b.size(0)
 
@@ -234,7 +236,7 @@ def batch_l1_proj_flat(x, z=1):
     view_size = view.size(1)
     mu = view.abs().sort(1, descending=True)[0]
     vv = torch.arange(view_size).float().to(x.device)
-    st = (mu.cumsum(1) - z[indexes_b][:, None]) / (vv + 1)
+    st = (mu.cumsum(1) - z) / (vv + 1)
     u = (mu - st) > 0
     if u.dtype.__str__() == "torch.bool":  # after and including torch 1.2
         rho = (~u).cumsum(dim=1).eq(0).sum(1) - 1

docs/conf.py

@@ -23,7 +23,7 @@
 os.symlink(
     "../../advertorch_examples/tutorial_attack_defense_bpda_mnist.ipynb",
     "_tutorials/tutorial_attack_defense_bpda_mnist.ipynb")
-import sys  # noqa: F401
+import sys  # noqa: F401, E402
 sys.path.insert(0, os.path.abspath('..'))
 
 
@@ -52,7 +52,7 @@
 # 'scipy._lib',
 # ]
 
-from unittest.mock import Mock  # noqa: F401
+from unittest.mock import Mock  # noqa: F401, E402
 # from sphinx.ext.autodoc.importer import _MockObject as Mock
 Mock.Module = object
 sys.modules['torch'] = Mock()
@@ -85,7 +85,7 @@
 sys.modules['scipy._lib'] = Mock()
 
 # XXX: This import has to be after mock
-import advertorch  # noqa: F401
+import advertorch  # noqa: F401, E402
 
 
 # -- Project information -----------------------------------------------------

tests/test_attacks_running.py

@@ -17,11 +17,12 @@
 import torch.nn as nn
 
 from advertorch.attacks import GradientSignAttack
-from advertorch.attacks import LinfBasicIterativeAttack
 from advertorch.attacks import GradientAttack
 from advertorch.attacks import L2BasicIterativeAttack
-from advertorch.attacks import LinfPGDAttack
+from advertorch.attacks import LinfBasicIterativeAttack
 from advertorch.attacks import L1PGDAttack
+from advertorch.attacks import L2PGDAttack
+from advertorch.attacks import LinfPGDAttack
 from advertorch.attacks import SparseL1DescentAttack
 from advertorch.attacks import MomentumIterativeAttack
 from advertorch.attacks import FastFeatureAttack
@@ -71,9 +72,15 @@
 attack_kwargs = {
     GradientSignAttack: {},
     GradientAttack: {},
-    LinfBasicIterativeAttack: {"nb_iter": 5},
-    L2BasicIterativeAttack: {"nb_iter": 5},
-    LinfPGDAttack: {"rand_init": False, "nb_iter": 5},
+    SparseL1DescentAttack: {
+        "rand_init": False, "nb_iter": 5, "eps": 3., "eps_iter": 1.},
+    L1PGDAttack: {"rand_init": False, "nb_iter": 5, "eps": 3., "eps_iter": 1.},
+    L2BasicIterativeAttack: {"nb_iter": 5, "eps": 1., "eps_iter": 0.33},
+    L2PGDAttack: {
+        "rand_init": False, "nb_iter": 5, "eps": 1., "eps_iter": 0.33},
+    LinfBasicIterativeAttack: {"nb_iter": 5, "eps": 0.3, "eps_iter": 0.1},
+    LinfPGDAttack: {
+        "rand_init": False, "nb_iter": 5, "eps": 0.3, "eps_iter": 0.1},
     MomentumIterativeAttack: {"nb_iter": 5},
     CarliniWagnerL2Attack: {"num_classes": NUM_CLASS, "max_iterations": 10},
     ElasticNetL1Attack: {"num_classes": NUM_CLASS, "max_iterations": 10},
@@ -82,8 +89,6 @@
     JacobianSaliencyMapAttack: {"num_classes": NUM_CLASS, "gamma": 0.01},
     SpatialTransformAttack: {"num_classes": NUM_CLASS},
     DDNL2Attack: {"nb_iter": 5},
-    SparseL1DescentAttack: {"rand_init": False, "nb_iter": 5},
-    L1PGDAttack: {"rand_init": False, "nb_iter": 5},
     LinfSPSAAttack: {"eps": 0.3, "max_batch_size": 63},
     LinfFABAttack: {"n_iter": 5},
     L2FABAttack: {"n_iter": 5},

tests/test_defenses_running.py

@@ -64,7 +64,7 @@ def test_withgrad(device, def_cls):
     "device, def_cls",
     itertools.product(devices, nograd_defenses))
 def test_defenses_nograd(device, def_cls):
-    with pytest.raises(NotImplementedError):
+    with pytest.raises((RuntimeError, NotImplementedError)):
         defense = def_cls(**defense_kwargs[def_cls])
         data = defense_data[def_cls]
         data.requires_grad_()