x_test, y_test, inp_shape, num_classes = load_dataset(
dataset_name=m["dataset"], shuffle=False, n_inputs=n_inputs)[2:]
det_model_savedir = get_model_savedir(model="baseNN",
dataset=m["dataset"],
architecture=m["architecture"],
debug=args.debug,
model_idx=args.model_idx)
detnet = baseNN(inp_shape, num_classes, *list(m.values()))
detnet.load(savedir=det_model_savedir, device=args.device)
det_attacks = load_attack(method=args.attack_method,
model_savedir=det_model_savedir)
det_predictions, det_atk_predictions, det_softmax_robustness, det_successful_idxs, det_failed_idxs = \
evaluate_attack(net=detnet, x_test=x_test, x_attack=det_attacks, y_test=y_test,
device=args.device, return_classification_idxs=True)
m = fullBNN_settings["model_" + str(args.model_idx)]
x_test, y_test, inp_shape, out_size = load_dataset(dataset_name=m["dataset"],
shuffle=False,
n_inputs=n_inputs)[2:]
bay_model_savedir = get_model_savedir(model="fullBNN",
dataset=m["dataset"],
architecture=m["architecture"],
model_idx=args.model_idx,
debug=args.debug)
bayesnet = BNN(m["dataset"], *list(m.values())[1:], inp_shape, num_classes)
bayesnet.load(savedir=bay_model_savedir, device=args.device)
model_savedir=savedir)
else:
x_attack = attack(net=net,
x_test=x_test,
y_test=y_test,
device=args.device,
method=args.attack_method)
save_attack(x_test,
x_attack,
method=args.attack_method,
model_savedir=savedir)
evaluate_attack(net=net,
x_test=x_test,
x_attack=x_attack,
y_test=y_test,
device=args.device)
else:
if args.model == "fullBNN":
m = fullBNN_settings["model_" + str(args.model_idx)]
x_test, y_test, inp_shape, out_size = load_dataset(
dataset_name=m["dataset"], n_inputs=n_inputs)[2:]
savedir = get_model_savedir(model=args.model,
dataset=m["dataset"],
architecture=m["architecture"],
for im_idx in range(det_lrp.shape[0]):
det_lrp[im_idx] = normalize(det_lrp[im_idx])
det_attack_lrp[im_idx] = normalize(det_attack_lrp[im_idx])
for samp_idx in range(len(n_samples_list)):
bay_lrp[samp_idx][im_idx] = normalize(
bay_lrp[samp_idx][im_idx])
bay_attack_lrp[samp_idx][im_idx] = normalize(
bay_attack_lrp[samp_idx][im_idx])
### Evaluate explanations
det_preds, det_atk_preds, det_softmax_robustness, det_successful_idxs, det_failed_idxs = evaluate_attack(
net=detnet,
x_test=images,
x_attack=det_attack,
y_test=y_test,
device=args.device,
return_classification_idxs=True)
det_softmax_robustness = det_softmax_robustness.detach().cpu().numpy()
det_lrp_robustness, det_lrp_pxl_idxs = lrp_robustness(
original_heatmaps=det_lrp,
adversarial_heatmaps=det_attack_lrp,
topk=topk,
method=lrp_robustness_method)
succ_det_lrp_robustness, succ_det_lrp_pxl_idxs = lrp_robustness(
original_heatmaps=det_lrp[det_successful_idxs],
adversarial_heatmaps=det_attack_lrp[det_successful_idxs],
topk=topk,
dataset_name=m["dataset"], shuffle=False, n_inputs=n_inputs)[2:]
model_savedir = get_model_savedir(model="baseNN",
dataset=m["dataset"],
architecture=m["architecture"],
debug=args.debug,
model_idx=args.model_idx)
detnet = baseNN(inp_shape, num_classes, *list(m.values()))
detnet.load(savedir=model_savedir, device=args.device)
attacks = load_attack(method=args.attack_method,
model_savedir=model_savedir)
predictions, atk_predictions, softmax_robustness, successful_idxs, failed_idxs = evaluate_attack(
net=detnet,
x_test=x_test,
x_attack=attacks,
y_test=y_test,
device=args.device,
return_classification_idxs=True)
learnable_layers_idxs = detnet.learnable_layers_idxs
elif args.model == "fullBNN":
m = fullBNN_settings["model_" + str(args.model_idx)]
x_test, y_test, inp_shape, num_classes = load_dataset(
dataset_name=m["dataset"], shuffle=False, n_inputs=n_inputs)[2:]
model_savedir = get_model_savedir(model=args.model,
dataset=m["dataset"],
architecture=m["architecture"],