def run_restrictedImgnetHypo(auto_var):
device = 'cuda' if torch.cuda.is_available() else 'cpu'
_ = set_random_seed(auto_var)
norm = auto_var.get_var("norm")
trn_ds, tst_ds = auto_var.get_var("dataset")
lbl_enc = None # TODO
auto_var.set_intermidiate_variable("lbl_enc", lbl_enc)
n_classes = len(trn_ds.classes)
result = {}
mock_trnX = np.concatenate([trn_ds[0][0], trn_ds[1][0]], axis=0)
trny = np.array(trn_ds.targets)
tsty = np.array(tst_ds.targets)
#multigpu = True if torch.cuda.device_count() > 1 else False
multigpu = False
model = auto_var.get_var("model", trnX=trnX, trny=trny, multigpu=multigpu, n_channels=n_channels)
model.tst_ds = (tstX, tsty)
with Stopwatch("Fitting Model"):
history = model.fit(trnX, trny)
result['trn_acc'] = (model.predict_ds(trn_ds) == trny).mean()
result['tst_acc'] = (model.predict_ds(tst_ds) == tsty).mean()
print(f"train acc: {result['trn_acc']}")
print(f"test acc: {result['tst_acc']}")
attack_model = auto_var.get_var("attack", model=model, n_classes=n_classes,
clip_min=0, clip_max=1,)
with Stopwatch("Attacking Train"):
adv_trnX = attack_model.perturb_ds(trn_ds)
with Stopwatch("Attacking Test"):
adv_tstX = attack_model.perturb_ds(tst_ds)
result['adv_trn_acc'] = (model.predict(adv_trnX) == trny).mean()
result['adv_tst_acc'] = (model.predict(adv_tstX) == tsty).mean()
print(f"adv trn acc: {result['adv_trn_acc']}")
print(f"adv tst acc: {result['adv_tst_acc']}")
del attack_model
with Stopwatch("Estimating trn Lip"):
trn_lip, _ = estimate_local_lip_v2(model.model, trn_ds, top_norm=1, btm_norm=norm,
epsilon=auto_var.get_var("eps"), device=device)
result['avg_trn_lip_1'] = trn_lip
with Stopwatch("Estimating tst Lip"):
tst_lip, _ = estimate_local_lip_v2(model.model, tst_ds, top_norm=1, btm_norm=norm,
epsilon=auto_var.get_var("eps"), device=device)
result['avg_tst_lip_1'] = tst_lip
print(f"avg trn lip: {result['avg_trn_lip_1']}")
print(f"avg tst lip: {result['avg_tst_lip_1']}")
print(result)
return result
def run_hypo(auto_var):
device = 'cuda' if torch.cuda.is_available() else 'cpu'
_ = set_random_seed(auto_var)
norm = auto_var.get_var("norm")
trnX, trny, tstX, tsty = auto_var.get_var("dataset")
lbl_enc = OneHotEncoder(categories=[np.sort(np.unique(trny))],
sparse=False).fit(trny.reshape(-1, 1))
auto_var.set_intermidiate_variable("lbl_enc", lbl_enc)
n_classes = len(np.unique(trny))
n_channels = trnX.shape[-1]
result = {}
multigpu = False
model = auto_var.get_var("model",
trnX=trnX,
trny=trny,
multigpu=multigpu,
n_channels=n_channels)
model.tst_ds = (tstX, tsty)
with Stopwatch("Fitting Model"):
history = model.fit(trnX, trny)
result['trn_acc'] = (model.predict(trnX) == trny).mean()
result['tst_acc'] = (model.predict(tstX) == tsty).mean()
print(f"train acc: {result['trn_acc']}")
print(f"test acc: {result['tst_acc']}")
attack_model = auto_var.get_var(
"attack",
model=model,
n_classes=n_classes,
clip_min=0,
clip_max=1,
)
with Stopwatch("Attacking"):
adv_trnX = attack_model.perturb(trnX, trny)
adv_tstX = attack_model.perturb(tstX, tsty)
result['adv_trn_acc'] = (model.predict(adv_trnX) == trny).mean()
result['adv_tst_acc'] = (model.predict(adv_tstX) == tsty).mean()
print(f"adv trn acc: {result['adv_trn_acc']}")
print(f"adv tst acc: {result['adv_tst_acc']}")
del attack_model
with Stopwatch("Estimating trn Lip"):
trn_lip, _ = estimate_local_lip_v2(model.model,
trnX,
top_norm=1,
btm_norm=norm,
epsilon=auto_var.get_var("eps"),
device=device)
result['avg_trn_lip_1'] = trn_lip
with Stopwatch("Estimating tst Lip"):
tst_lip, _ = estimate_local_lip_v2(model.model,
tstX,
top_norm=1,
btm_norm=norm,
epsilon=auto_var.get_var("eps"),
device=device)
result['avg_tst_lip_1'] = tst_lip
print(f"avg trn lip: {result['avg_trn_lip_1']}")
print(f"avg tst lip: {result['avg_tst_lip_1']}")
print(result)
return result