def main():
ae_name = 'aae'
for dataset in ['mnist', 'fashion', 'cifar10']:
path = './'
path_aemodels = path + 'aemodels/%s/%s/' % (dataset, ae_name)
X_train, _, X_test, _, use_rgb = get_dataset(dataset)
ae = get_autoencoder(X_test, ae_name, dataset, path_aemodels)
ae.load_model()
X_train_ae = ae.decode(ae.encode(X_train))
X_test_ae = ae.decode(ae.encode(X_test))
print('dataset: ', dataset)
print('train rmse:', rmse(X_train, X_train_ae))
print('test rmse:', rmse(X_test, X_test_ae))
print('')
def main():
dataset = 'mnist'
ae_name = 'aae'
epochs = 10000
batch_size = 256
sample_interval = 200
path = './'
path_aemodels = path + 'aemodels/%s/%s/' % (dataset, ae_name)
_, _, X_test, Y_test, use_rgb = get_dataset(dataset)
ae = get_autoencoder(X_test, ae_name, dataset, path_aemodels)
ae.fit(X_test,
epochs=epochs,
batch_size=batch_size,
sample_interval=sample_interval)
ae.save_model()
ae.sample_images(epochs)
def main():
dataset = sys.argv[1]
black_box = 'RF'
neigh_type = 'hrgp'
random_state = 0
ae_name = 'aae'
num_classes = 10
nbr_experiments = 200
if dataset not in ['mnist', 'cifar10', 'fashion']:
print('unknown dataset %s' % dataset)
return -1
if black_box not in ['RF', 'AB', 'DNN']:
print('unknown black box %s' % black_box)
return -1
if neigh_type not in ['rnd', 'gntp', 'hrgp']:
print('unknown neigh type %s' % neigh_type)
return -1
path = './'
path_models = path + 'models/'
path_results = path + 'results/rel/'
path_aemodels = path + 'aemodels/%s/%s/' % (dataset, ae_name)
path_expl = './expl/'
black_box_filename = path_models + '%s_%s' % (dataset, black_box)
results_filename = path_results + 'rel_%s_%s_%s.json' % (dataset, black_box, neigh_type)
expl_filename = path_expl + 'alore_%s_%s_%s.json.gz' % (dataset, black_box, neigh_type)
_, _, X_test, Y_test, use_rgb = get_dataset(dataset)
bb, transform = get_black_box(black_box, black_box_filename, use_rgb, return_model=True)
bb_predict, bb_predict_proba = get_black_box(black_box, black_box_filename, use_rgb)
ae = get_autoencoder(X_test, ae_name, dataset, path_aemodels)
ae.load_model()
class_name = 'class'
class_values = ['%s' % i for i in range(len(np.unique(Y_test)))]
explainer = ILOREM(bb_predict, class_name, class_values, neigh_type=neigh_type, use_prob=True, size=1000, ocr=0.1,
kernel_width=None, kernel=None, autoencoder=ae, use_rgb=use_rgb, valid_thr=0.5,
filter_crules=True, random_state=random_state, verbose=False, alpha1=0.5, alpha2=0.5,
metric=neuclidean, ngen=10, mutpb=0.2, cxpb=0.5, tournsize=3, halloffame_ratio=0.1,
bb_predict_proba=bb_predict_proba)
errors = open(path_results + 'errors_relevancy_%s_%s.csv' % (dataset, black_box), 'w')
for i2e in range(nbr_experiments):
jrow_rel_o = {'i2e': i2e, 'dataset': dataset, 'black_box': black_box}
img = X_test[i2e]
bbo = bb_predict(np.array([img]))
jrow_list = list()
try:
# Alore
exp = explainer.explain_instance(img, num_samples=1000, use_weights=True, metric=neuclidean)
_, diff = exp.get_image_rule(features=None, samples=100)
relevancy = 1.0 - np.abs(diff - 127.5)/127.5
for color in [0, 127, 255]:
jrow_rel = copy.deepcopy(jrow_rel_o)
jrow_rel['method'] = 'alore'
jrow_rel['color'] = color
jrow_rel = apply_relevancy(jrow_rel, img, bb_predict, bbo[0], relevancy, color)
jrow_list.append(jrow_rel)
print(datetime.datetime.now(),
'[%s/%s] %s %s %s %s - 25: %d, 50: %d, 75: %d' % (i2e, nbr_experiments, dataset, black_box,
'alore', color, jrow_rel['color%s_c25' % color],
jrow_rel['color%s_c50' % color], jrow_rel['color%s_c75' % color]))
jrow_neigh = {'i2e': i2e, 'dataset': dataset, 'black_box': black_box, 'expl': diff,
'rule': str(exp.rule)}
json_str = ('%s\n' % json.dumps(jrow_neigh, cls=NumpyEncoder)).encode('utf-8')
with gzip.GzipFile(expl_filename, 'a') as fout:
fout.write(json_str)
except Exception:
print('error instance to explain: %d' % i2e)
errors.write('%d\n' % i2e)
continue
results = open(results_filename, 'a')
for jrow in jrow_list:
results.write('%s\n' % json.dumps(jrow))
results.close()
errors.close()
def main():
dataset = sys.argv[1]
black_box = 'DNN'
neigh_type = 'hrgp'
if len(sys.argv) > 2:
start_from = int(sys.argv[2])
else:
start_from = 0
random_state = 0
ae_name = 'aae'
num_classes = 10
nbr_experiments = 200
if dataset not in ['mnist', 'cifar10', 'fashion']:
print('unknown dataset %s' % dataset)
return -1
if black_box not in ['RF', 'AB', 'DNN']:
print('unknown black box %s' % black_box)
return -1
if neigh_type not in ['rnd', 'gntp', 'hrgp']:
print('unknown neigh type %s' % neigh_type)
return -1
path = './'
path_models = path + 'models/'
path_results = path + 'results/stability/'
path_aemodels = path + 'aemodels/%s/%s/' % (dataset, ae_name)
black_box_filename = path_models + '%s_%s' % (dataset, black_box)
results_filename = path_results + 'sta_%s_%s_%s.json' % (
dataset, black_box, neigh_type)
_, _, X_test, Y_test, use_rgb = get_dataset(dataset)
bb, transform = get_black_box(black_box,
black_box_filename,
use_rgb,
return_model=True)
bb_predict, bb_predict_proba = get_black_box(black_box, black_box_filename,
use_rgb)
Y_pred = bb_predict(X_test)
Y_pred_proba = bb_predict_proba(X_test)
ae = get_autoencoder(X_test, ae_name, dataset, path_aemodels)
ae.load_model()
class_name = 'class'
class_values = ['%s' % i for i in range(len(np.unique(Y_test)))]
explainer = ILOREM(bb_predict,
class_name,
class_values,
neigh_type=neigh_type,
use_prob=True,
size=1000,
ocr=0.1,
kernel_width=None,
kernel=None,
autoencoder=ae,
use_rgb=use_rgb,
valid_thr=0.5,
filter_crules=True,
random_state=random_state,
verbose=False,
alpha1=0.5,
alpha2=0.5,
metric=neuclidean,
ngen=10,
mutpb=0.2,
cxpb=0.5,
tournsize=3,
halloffame_ratio=0.1,
bb_predict_proba=bb_predict_proba)
lime_explainer = lime_image.LimeImageExplainer()
segmenter = SegmentationAlgorithm('quickshift',
kernel_size=1,
max_dist=200,
ratio=0.2)
input_tensor = bb.layers[0].input
last_layer = -2 if dataset == 'mnist' else -1
bb_model = Model(inputs=input_tensor, outputs=bb.layers[last_layer].output)
target_tensor = bb_model(input_tensor)
de_list = ['grad*input', 'saliency', 'intgrad', 'elrp', 'occlusion']
errors = open(
path_results + 'errors_stability_%s_%s.csv' % (dataset, black_box),
'w')
with DeepExplain(session=K.get_session()) as de:
for i2e in range(nbr_experiments):
if i2e < start_from:
continue
try:
print(
datetime.datetime.now(),
'[%s/%s] %s %s - checking stability' %
(i2e, nbr_experiments, dataset, black_box))
expl_list = list()
jrow_list = list()
jrow_coh_o = {
'i2e': i2e,
'dataset': dataset,
'black_box': black_box
}
# Crate random noise
img = X_test[i2e]
bbo = bb_predict(np.array([img]))
bbop = Y_pred_proba[i2e]
X_random_noise = generate_random_noise(img,
bb_predict,
bbo[0],
nbr_samples=20)
# Y_pred_random_noise = bb_predict(X_random_noise)
Y_pred_proba_random_noise = bb_predict_proba(X_random_noise)
# plt.subplot(1, 3, 1)
# plt.imshow(X_random_noise[0], cmap='gray')
# plt.subplot(1, 3, 2)
# plt.imshow(X_random_noise[1], cmap='gray')
# plt.subplot(1, 3, 3)
# plt.imshow(X_random_noise[2], cmap='gray')
# plt.show()
# Alore
print(datetime.datetime.now(), 'calculating alore')
exp = explainer.explain_instance(img,
num_samples=1000,
use_weights=True,
metric=neuclidean)
_, diff = exp.get_image_rule(features=None, samples=100)
expl_list.append(diff)
# Lime
print(datetime.datetime.now(), 'calculating lime')
exp = lime_explainer.explain_instance(
img,
bb_predict_proba,
top_labels=1,
hide_color=0,
num_samples=1000,
segmentation_fn=segmenter)
_, mask = exp.get_image_and_mask(bbo[0],
positive_only=False,
num_features=5,
hide_rest=False,
min_weight=0.01)
expl_list.append(mask)
# Deep Explain
xs = transform(np.array([img]))
ys = to_categorical(bbo, num_classes)
for det in de_list:
print(datetime.datetime.now(), 'calculating %s' % det)
if det == 'shapley_sampling':
maps = de.explain(det,
target_tensor,
input_tensor,
xs,
ys=ys,
samples=10)[0]
else:
maps = de.explain(det,
target_tensor,
input_tensor,
xs,
ys=ys)[0]
maps = np.mean(maps, axis=2)
expl_list.append(maps)
lipschitz_list = defaultdict(list)
lipschitz_list_bb = defaultdict(list)
print(datetime.datetime.now(), 'calculating lipschitz')
for i2e1 in range(len(X_random_noise)):
img1 = X_random_noise[i2e1]
bbo1 = bb_predict(np.array([img1]))
bbop1 = Y_pred_proba_random_noise[i2e1]
norm_bb = calculate_lipschitz_factor(bbop, bbop1)
norm_x = calculate_lipschitz_factor(img, img1)
# Alore
exp1 = explainer.explain_instance(img1,
num_samples=1000,
use_weights=True,
metric=neuclidean)
_, diff1 = exp1.get_image_rule(features=None, samples=100)
norm_exp = calculate_lipschitz_factor(expl_list[0], diff1)
lipschitz_list['alore'].append(norm_exp / norm_x)
lipschitz_list_bb['alore'].append(norm_exp / norm_bb)
print(datetime.datetime.now(), '\talore',
norm_exp / norm_x)
# Lime
exp1 = lime_explainer.explain_instance(
img1,
bb_predict_proba,
top_labels=1,
hide_color=0,
num_samples=1000,
segmentation_fn=segmenter)
_, mask1 = exp1.get_image_and_mask(bbo[0],
positive_only=False,
num_features=5,
hide_rest=False,
min_weight=0.01)
norm_exp = calculate_lipschitz_factor(expl_list[1], mask1)
lipschitz_list['lime'].append(norm_exp / norm_x)
lipschitz_list_bb['lime'].append(norm_exp / norm_bb)
print(datetime.datetime.now(), '\tlime', norm_exp / norm_x)
# DeepExplain
xs1 = transform(np.array([img1]))
ys1 = to_categorical(bbo1, num_classes)
for i, det in enumerate(de_list):
if det == 'shapley_sampling':
maps1 = de.explain(det,
target_tensor,
input_tensor,
xs1,
ys=ys1,
samples=10)[0]
else:
maps1 = de.explain(det,
target_tensor,
input_tensor,
xs1,
ys=ys1)[0]
maps1 = np.mean(maps1, axis=2)
norm_exp = calculate_lipschitz_factor(
expl_list[i + 2], maps1)
lipschitz_list[det].append(norm_exp / norm_x)
lipschitz_list_bb[det].append(norm_exp / norm_bb)
print(datetime.datetime.now(), '\t%s' % det,
norm_exp / norm_x)
for k in lipschitz_list:
jrow_coh = copy.deepcopy(jrow_coh_o)
jrow_coh['method'] = k
jrow_coh['mean'] = float(np.nanmean(lipschitz_list[k]))
jrow_coh['std'] = float(np.nanstd(lipschitz_list[k]))
jrow_coh['max'] = float(np.nanmax(lipschitz_list[k]))
jrow_coh['mean_bb'] = float(
np.nanmean(lipschitz_list_bb[k]))
jrow_coh['std_bb'] = float(np.nanstd(lipschitz_list_bb[k]))
jrow_coh['max_bb'] = float(np.nanmax(lipschitz_list_bb[k]))
jrow_list.append(jrow_coh)
print(
datetime.datetime.now(),
'[%s/%s] %s %s %s - mean: %.3f, max: %.3f' %
(i2e, nbr_experiments, dataset, black_box, k,
jrow_coh['mean'], jrow_coh['max']))
except Exception:
print('error instance to explain: %d' % i2e)
errors.write('%d\n' % i2e)
continue
results = open(results_filename, 'a')
for jrow in jrow_list:
results.write('%s\n' % json.dumps(jrow))
results.close()
errors.close()
def main():
dataset = sys.argv[1]
black_box = sys.argv[2]
neigh_type = sys.argv[3]
# dataset = 'mnist'
# black_box = 'DNN'
# neigh_type = 'hrgp'
max_nbr_exemplars = 128
random_state = 0
ae_name = 'aae'
nbr_experiments = 200
if dataset not in ['mnist', 'cifar10', 'fashion']:
print('unknown dataset %s' % dataset)
return -1
if black_box not in ['RF', 'AB', 'DNN']:
print('unknown black box %s' % black_box)
return -1
if neigh_type not in ['rnd', 'gntp', 'hrgp']:
print('unknown neigh type %s' % neigh_type)
return -1
path = './'
path_models = path + 'models/'
path_results = path + 'results/validity/'
path_aemodels = path + 'aemodels/%s/%s/' % (dataset, ae_name)
black_box_filename = path_models + '%s_%s' % (dataset, black_box)
results_filename = path_results + 'validthr_%s_%s_%s.json' % (
dataset, black_box, neigh_type)
_, _, X_test, Y_test, use_rgb = get_dataset(dataset)
bb, transform = get_black_box(black_box,
black_box_filename,
use_rgb,
return_model=True)
bb_predict, bb_predict_proba = get_black_box(black_box, black_box_filename,
use_rgb)
ae = get_autoencoder(X_test, ae_name, dataset, path_aemodels)
ae.load_model()
Y_pred = bb_predict(X_test)
Y_pred_proba = bb_predict_proba(X_test)
X_test_comp = X_test[nbr_experiments:]
Y_pred_comp = Y_pred[nbr_experiments:]
Y_pred_proba_comp = Y_pred_proba[nbr_experiments:]
class_name = 'class'
class_values = ['%s' % i for i in range(len(np.unique(Y_test)))]
errors = open(
path_results + 'errors_validity_%s_%s_%s.csv' %
(dataset, black_box, neigh_type), 'w')
for i2e in range(nbr_experiments):
print(datetime.datetime.now(),
'[%s/%s] %s %s' % (i2e, nbr_experiments, dataset, black_box))
for valid_thr in np.arange(0.0, 1.0, 0.1):
explainer = ILOREM(bb_predict,
class_name,
class_values,
neigh_type=neigh_type,
use_prob=True,
size=1000,
ocr=0.1,
kernel_width=None,
kernel=None,
autoencoder=ae,
use_rgb=use_rgb,
valid_thr=valid_thr,
filter_crules=True,
random_state=random_state,
verbose=False,
alpha1=0.5,
alpha2=0.5,
metric=neuclidean,
ngen=10,
mutpb=0.2,
cxpb=0.5,
tournsize=3,
halloffame_ratio=0.1,
bb_predict_proba=bb_predict_proba)
try:
jrow_o = {
'i2e': i2e,
'dataset': dataset,
'black_box': black_box,
'neigh_type': neigh_type,
'valid_thr': valid_thr
}
jrow_list = list()
img = X_test[i2e]
bbo = bb_predict(np.array([img]))
start_time = datetime.datetime.now()
print(datetime.datetime.now(), '\textract explanation',
valid_thr)
exp = explainer.explain_instance(img,
num_samples=1000,
use_weights=True,
metric=neuclidean)
run_time = (datetime.datetime.now() -
start_time).total_seconds()
jrow_o['time'] = run_time
fidelity = exp.fidelity
jrow_o['fidelity'] = fidelity
print(datetime.datetime.now(), '\tcalculate compactness',
valid_thr)
res_compactness = get_compactness(img, exp, black_box,
transform)
compact, compact_var, lcompact, lcompact_var, img_cdist, rdist = res_compactness
jrow_o['compactness'] = compact
jrow_o['compactness_var'] = compact_var
jrow_o['lcompactness'] = lcompact
jrow_o['lcompactness_var'] = lcompact_var
print(datetime.datetime.now(), '\tcalculate plausibility',
valid_thr)
plausibility = get_plausibility(img_cdist, black_box,
transform, X_test, rdist)
for i, p in enumerate(plausibility):
jrow_o['plausibility%d' % i] = p
print(datetime.datetime.now(), '\tcalculate saliency map',
valid_thr)
_, diff = exp.get_image_rule(features=None, samples=100)
print(datetime.datetime.now(), '\tcalculate relevancy',
valid_thr)
relevancy = 1.0 - np.abs(diff - 127.5) / 127.5
for color in [0, 127, 255]:
jrow_o = apply_relevancy(jrow_o, img, bb_predict, bbo[0],
relevancy, color)
print(datetime.datetime.now(), '\tcalculate coherence',
valid_thr)
coherence_lipschitz, coherence_lipschitz_bb = get_lipswhitz_coherence(
img, i2e, bb_predict, X_test_comp, Y_pred_comp,
Y_pred_proba, Y_pred_proba_comp, diff, explainer)
jrow_o['coherence_mean'] = float(
np.nanmean(coherence_lipschitz))
jrow_o['coherence_std'] = float(np.nanstd(coherence_lipschitz))
jrow_o['coherence_max'] = float(np.nanmax(coherence_lipschitz))
jrow_o['coherence_mean_bb'] = float(
np.nanmean(coherence_lipschitz_bb))
jrow_o['coherence_std_bb'] = float(
np.nanstd(coherence_lipschitz_bb))
jrow_o['coherence_max_bb'] = float(
np.nanmax(coherence_lipschitz_bb))
print(datetime.datetime.now(), '\tcalculate stability',
valid_thr)
stability_lipschitz, stability_lipschitz_bb = get_lipswhitz_stability(
img, i2e, bb_predict, X_test_comp, Y_pred_proba,
Y_pred_proba_comp, diff, explainer)
jrow_o['stability_mean'] = float(
np.nanmean(stability_lipschitz))
jrow_o['stability_std'] = float(np.nanstd(stability_lipschitz))
jrow_o['stability_max'] = float(np.nanmax(stability_lipschitz))
jrow_o['stability_mean_bb'] = float(
np.nanmean(stability_lipschitz_bb))
jrow_o['stability_std_bb'] = float(
np.nanstd(stability_lipschitz_bb))
jrow_o['stability_max_bb'] = float(
np.nanmax(stability_lipschitz_bb))
print(datetime.datetime.now(), '\tcalculate knn', valid_thr)
exemplars = exp.get_prototypes_respecting_rule(
num_prototypes=max_nbr_exemplars)
cexemplars = exp.get_counterfactual_prototypes(eps=0.01)
if len(cexemplars) < max_nbr_exemplars:
cexemplars2 = exp.get_prototypes_not_respecting_rule(
num_prototypes=max_nbr_exemplars - len(cexemplars))
cexemplars.extend(cexemplars2)
X_test_knn, Y_test_knn = prepare_test_for_knn(
bbo, X_test_comp, Y_pred_comp, bb_predict, 200)
for nbr_exemplars in [1, 2, 4, 8, 16, 32, 64, 128]:
jrow_e = copy.deepcopy(jrow_o)
jrow_e['nbr_exemplars'] = nbr_exemplars
X_train_knn, Y_train_knn = prepare_data_for_knn(
exemplars[:nbr_exemplars], cexemplars[:nbr_exemplars],
bb_predict(np.array(exemplars[:nbr_exemplars])),
bb_predict(np.array(cexemplars[:nbr_exemplars])))
for k in range(1, min(nbr_exemplars + 1, 11)):
acc = evaluate_with_knn(X_train_knn, Y_train_knn,
X_test_knn, Y_test_knn, k)
jrow = copy.deepcopy(jrow_e)
jrow['k'] = k
jrow['accuracy'] = acc
jrow_list.append(jrow)
results = open(results_filename, 'a')
for jrow in jrow_list:
results.write('%s\n' % json.dumps(jrow))
results.close()
except Exception:
print('error instance to explain: %d' % i2e)
errors.write('%d\n' % i2e)
continue
errors.close()
def main():
dataset = sys.argv[1]
# dataset = 'mnist'
if len(sys.argv) > 2:
start_from = int(sys.argv[2])
else:
start_from = 0
black_box = 'DNN'
neigh_type = 'hrgp'
max_nbr_exemplars = 128
random_state = 0
ae_name = 'aae'
gamma = None
nbr_experiments = 200
if dataset not in ['mnist', 'cifar10', 'fashion']:
print('unknown dataset %s' % dataset)
return -1
if black_box not in ['RF', 'AB', 'DNN']:
print('unknown black box %s' % black_box)
return -1
if neigh_type not in ['rnd', 'gntp', 'hrgp']:
print('unknown neigh type %s' % neigh_type)
return -1
path = './'
path_models = path + 'models/'
path_results = path + 'results/expcl/'
path_aemodels = path + 'aemodels/%s/%s/' % (dataset, ae_name)
black_box_filename = path_models + '%s_%s' % (dataset, black_box)
results_filename = path_results + 'expcl_%s_%s_%s.json' % (dataset, black_box, neigh_type)
_, _, X_test, Y_test, use_rgb = get_dataset(dataset)
bb, transform = get_black_box(black_box, black_box_filename, use_rgb, return_model=True)
bb_predict, bb_predict_proba = get_black_box(black_box, black_box_filename, use_rgb)
Y_pred = bb_predict(X_test)
X_test_comp = X_test[nbr_experiments:]
Y_pred_comp = Y_pred[nbr_experiments:]
X = np.array([x.ravel() for x in transform(X_test_comp)])
ae = get_autoencoder(X_test, ae_name, dataset, path_aemodels)
ae.load_model()
class_name = 'class'
class_values = ['%s' % i for i in range(len(np.unique(Y_test)))]
explainer = ILOREM(bb_predict, class_name, class_values, neigh_type=neigh_type, use_prob=True, size=1000, ocr=0.1,
kernel_width=None, kernel=None, autoencoder=ae, use_rgb=use_rgb, valid_thr=0.5,
filter_crules=True, random_state=random_state, verbose=False, alpha1=0.5, alpha2=0.5,
metric=neuclidean, ngen=10, mutpb=0.2, cxpb=0.5, tournsize=3, halloffame_ratio=0.1,
bb_predict_proba=bb_predict_proba)
errors = open(path_results + 'errors_expcl_%s_%s.csv' % (dataset, black_box), 'w')
for i2e in range(nbr_experiments):
if i2e < start_from:
continue
print(datetime.datetime.now(), '[%s/%s] %s %s' % (i2e, nbr_experiments, dataset, black_box))
try:
img = X_test[i2e]
bbo = bb_predict(np.array([img]))
jrow_o = {'i2e': i2e, 'dataset': dataset, 'black_box': black_box}
jrow_list = list()
X_test_knn, Y_test_knn = prepare_test_for_knn(bbo, X_test_comp, Y_pred_comp, bb_predict, 200)
# Alore
print(datetime.datetime.now(), 'alore')
exp = explainer.explain_instance(img, num_samples=1000, use_weights=True, metric=neuclidean)
# MMD
# kernel = rbf_kernel(X, gamma=gamma)
print(datetime.datetime.now(), 'mmd')
kernel = calculate_kernel_individual(X, Y_pred_comp, gamma)
# K-Medoids
print(datetime.datetime.now(), 'k-medoids')
X_idx = np.where(Y_pred_comp == bbo)[0]
Z = X[X_idx]
scaler = MinMaxScaler()
Z = scaler.fit_transform(Z)
zimg = scaler.transform(transform(np.array([img])).ravel().reshape(1, -1))
dist = cdist(zimg.reshape(1, -1), Z)
idx_p = np.argsort(dist)
idx_p = np.array([X_idx[i] for i in idx_p])
idx_p = idx_p[np.where(idx_p != i2e)]
X_idx = np.where(Y_pred_comp != bbo)[0]
Z = X[X_idx]
scaler = MinMaxScaler()
Z = scaler.fit_transform(Z)
zimg = scaler.transform(transform(np.array([img])).ravel().reshape(1, -1))
dist = cdist(zimg.reshape(1, -1), Z)
idx_n = np.argsort(dist)
idx_n = np.array([X_idx[i] for i in idx_n])
idx_n = idx_n[np.where(idx_n != i2e)]
# Alore
print(datetime.datetime.now(), 'alore - generate exemplars')
alore_exemplars = exp.get_prototypes_respecting_rule(num_prototypes=max_nbr_exemplars)
alore_cexemplars = exp.get_counterfactual_prototypes(eps=0.01)
if len(alore_cexemplars) < max_nbr_exemplars:
cexemplars2 = exp.get_prototypes_not_respecting_rule(num_prototypes=max_nbr_exemplars - len(alore_cexemplars))
alore_cexemplars.extend(cexemplars2)
# MMD
print(datetime.datetime.now(), 'mmd - select prototypes')
proto_idx = greedy_select_protos(kernel, np.array(range(np.shape(kernel)[0])),
nbr_prototypes=max_nbr_exemplars)
crit_idx = select_criticism_regularized(kernel, proto_idx, nbr_criticisms=max_nbr_exemplars,
is_K_sparse=False, reg='logdet')
proto_idx_sel = [i for i in proto_idx if Y_pred_comp[i] == bbo[0]][:max_nbr_exemplars]
crit_idx_sel = [i for i in crit_idx if Y_pred_comp[i] != bbo[0]][:max_nbr_exemplars]
mmd_exemplars = X_test_comp[proto_idx_sel]
mmd_cexemplars = X_test_comp[crit_idx_sel]
# K-Medoids
print(datetime.datetime.now(), 'k-medoids - select prototypes')
kmedoids_exemplars = X_test_comp[idx_p[:max_nbr_exemplars]]
kmedoids_cexemplars = X_test_comp[idx_n[:max_nbr_exemplars]]
for nbr_exemplars in [1, 2, 4, 8, 16, 32, 64, 128]:
jrow_e = copy.deepcopy(jrow_o)
jrow_e['nbr_exemplars'] = nbr_exemplars
X_train_knn, Y_train_knn = prepare_data_for_knn(alore_exemplars[:nbr_exemplars],
alore_cexemplars[:nbr_exemplars],
bb_predict(np.array(alore_exemplars[:nbr_exemplars])),
bb_predict(np.array(alore_cexemplars[:nbr_exemplars])))
for k in range(1, min(nbr_exemplars + 1, 11)):
jrow = copy.deepcopy(jrow_e)
acc = evaluate_with_knn(X_train_knn, Y_train_knn, X_test_knn, Y_test_knn, k)
jrow['method'] = 'alore'
jrow['k'] = k
jrow['accuracy'] = acc
jrow_list.append(jrow)
print(datetime.datetime.now(),
'[%s/%s] %s %s %s - alore: %.3f' % (i2e, nbr_experiments, dataset, black_box, k, acc))
X_train_knn, Y_train_knn = prepare_data_for_knn(mmd_exemplars[:nbr_exemplars],
mmd_cexemplars[:nbr_exemplars],
bb_predict(np.array(mmd_exemplars[:nbr_exemplars])),
bb_predict(np.array(mmd_cexemplars[:nbr_exemplars])))
for k in range(1, min(nbr_exemplars + 1, 11)):
jrow = copy.deepcopy(jrow_e)
acc = evaluate_with_knn(X_train_knn, Y_train_knn, X_test_knn, Y_test_knn, k)
jrow['method'] = 'mmd'
jrow['k'] = k
jrow['accuracy'] = acc
jrow_list.append(jrow)
print(datetime.datetime.now(),
'[%s/%s] %s %s %s - mmd: %.3f' % (i2e, nbr_experiments, dataset, black_box, k, acc))
X_train_knn, Y_train_knn = prepare_data_for_knn(kmedoids_exemplars[:nbr_exemplars],
kmedoids_cexemplars[:nbr_exemplars],
bb_predict(np.array(kmedoids_exemplars[:nbr_exemplars])),
bb_predict(np.array(kmedoids_cexemplars[:nbr_exemplars])))
for k in range(1, min(nbr_exemplars + 1, 11)):
jrow = copy.deepcopy(jrow_e)
acc = evaluate_with_knn(X_train_knn, Y_train_knn, X_test_knn, Y_test_knn, k)
jrow['method'] = 'k-medoids'
jrow['k'] = k
jrow['accuracy'] = acc
jrow_list.append(jrow)
print(datetime.datetime.now(),
'[%s/%s] %s %s %s - k-medoids: %.3f' % (i2e, nbr_experiments, dataset, black_box, k, acc))
except Exception:
print('error instance to explain: %d' % i2e)
errors.write('%d\n' % i2e)
continue
results = open(results_filename, 'a')
for jrow in jrow_list:
results.write('%s\n' % json.dumps(jrow))
results.close()
errors.close()
def main():
dataset = sys.argv[1]
# dataset = 'mnist'
black_box = 'RF'
neigh_type = 'hrgp'
random_state = 0
ae_name = 'aae'
num_classes = 10
nbr_experiments = 200
if dataset not in ['mnist', 'cifar10', 'fashion']:
print('unknown dataset %s' % dataset)
return -1
if black_box not in ['RF', 'AB', 'DNN']:
print('unknown black box %s' % black_box)
return -1
if neigh_type not in ['rnd', 'gntp', 'hrgp']:
print('unknown neigh type %s' % neigh_type)
return -1
path = './'
path_models = path + 'models/'
path_results = path + 'results/coherence/'
path_aemodels = path + 'aemodels/%s/%s/' % (dataset, ae_name)
black_box_filename = path_models + '%s_%s' % (dataset, black_box)
results_filename = path_results + 'coh_%s_%s_%s.json' % (
dataset, black_box, neigh_type)
_, _, X_test, Y_test, use_rgb = get_dataset(dataset)
bb, transform = get_black_box(black_box,
black_box_filename,
use_rgb,
return_model=True)
bb_predict, bb_predict_proba = get_black_box(black_box, black_box_filename,
use_rgb)
Y_pred = bb_predict(X_test)
Y_pred_proba = bb_predict_proba(X_test)
X_test_comp = X_test[nbr_experiments:]
Y_pred_comp = Y_pred[nbr_experiments:]
Y_pred_proba_comp = Y_pred_proba[nbr_experiments:]
ae = get_autoencoder(X_test, ae_name, dataset, path_aemodels)
ae.load_model()
class_name = 'class'
class_values = ['%s' % i for i in range(len(np.unique(Y_test)))]
explainer = ILOREM(bb_predict,
class_name,
class_values,
neigh_type=neigh_type,
use_prob=True,
size=1000,
ocr=0.1,
kernel_width=None,
kernel=None,
autoencoder=ae,
use_rgb=use_rgb,
valid_thr=0.5,
filter_crules=True,
random_state=random_state,
verbose=False,
alpha1=0.5,
alpha2=0.5,
metric=neuclidean,
ngen=10,
mutpb=0.2,
cxpb=0.5,
tournsize=3,
halloffame_ratio=0.1,
bb_predict_proba=bb_predict_proba)
lime_explainer = lime_image.LimeImageExplainer()
segmenter = SegmentationAlgorithm('quickshift',
kernel_size=1,
max_dist=200,
ratio=0.2)
errors = open(
path_results + 'errors_coehrence_%s_%s.csv' % (dataset, black_box),
'w')
for i2e in range(nbr_experiments):
try:
expl_list = list()
jrow_list = list()
jrow_coh_o = {
'i2e': i2e,
'dataset': dataset,
'black_box': black_box
}
# Finding Lipswhitz neighborhood
img = X_test[i2e]
bbo = bb_predict(np.array([img]))
bbop = Y_pred_proba[i2e]
X_idx = np.where(Y_pred_comp == bbo[0])[0]
scaler = MinMaxScaler()
x0 = scaler.fit_transform(img.ravel().reshape(-1, 1))
Xj = scaler.fit_transform([x.ravel() for x in X_test_comp[X_idx]])
dist = cdist(x0.reshape(1, -1), Xj)[0]
eps = np.percentile(dist, 5)
X_idx_eps = X_idx[np.where(dist <= eps)]
# Alore
exp = explainer.explain_instance(img,
num_samples=1000,
use_weights=True,
metric=neuclidean)
_, diff = exp.get_image_rule(features=None, samples=100)
expl_list.append(diff)
# Lime
exp = lime_explainer.explain_instance(img,
bb_predict_proba,
top_labels=1,
hide_color=0,
num_samples=1000,
segmentation_fn=segmenter)
_, mask = exp.get_image_and_mask(bbo[0],
positive_only=False,
num_features=5,
hide_rest=False,
min_weight=0.01)
expl_list.append(mask)
lipschitz_list = defaultdict(list)
lipschitz_list_bb = defaultdict(list)
print(
datetime.datetime.now(), '[%s/%s] %s %s - checking coherence' %
(i2e, nbr_experiments, dataset, black_box))
for i2e1 in X_idx_eps[:20]:
img1 = X_test_comp[i2e1]
bbo1 = bb_predict(np.array([img1]))
bbop1 = Y_pred_proba_comp[i2e1]
norm_bb = calculate_lipschitz_factor(bbop, bbop1)
norm_x = calculate_lipschitz_factor(img, img1)
# Alore
exp1 = explainer.explain_instance(img1,
num_samples=1000,
use_weights=True,
metric=neuclidean)
_, diff1 = exp1.get_image_rule(features=None, samples=100)
norm_exp = calculate_lipschitz_factor(expl_list[0], diff1)
lipschitz_list['alore'].append(norm_exp / norm_x)
lipschitz_list_bb['alore'].append(norm_exp / norm_bb)
print(datetime.datetime.now(), '\talore', norm_exp / norm_x)
# Lime
exp1 = lime_explainer.explain_instance(
img1,
bb_predict_proba,
top_labels=1,
hide_color=0,
num_samples=1000,
segmentation_fn=segmenter)
_, mask1 = exp1.get_image_and_mask(bbo[0],
positive_only=False,
num_features=5,
hide_rest=False,
min_weight=0.01)
norm_exp = calculate_lipschitz_factor(expl_list[1], mask1)
lipschitz_list['lime'].append(norm_exp / norm_x)
lipschitz_list_bb['lime'].append(norm_exp / norm_bb)
print(datetime.datetime.now(), '\tlime', norm_exp / norm_x)
for k in lipschitz_list:
jrow_coh = copy.deepcopy(jrow_coh_o)
jrow_coh['method'] = k
jrow_coh['mean'] = float(np.nanmean(lipschitz_list[k]))
jrow_coh['std'] = float(np.nanstd(lipschitz_list[k]))
jrow_coh['max'] = float(np.nanmax(lipschitz_list[k]))
jrow_coh['mean_bb'] = float(np.nanmean(lipschitz_list_bb[k]))
jrow_coh['std_bb'] = float(np.nanstd(lipschitz_list_bb[k]))
jrow_coh['max_bb'] = float(np.nanmax(lipschitz_list_bb[k]))
jrow_list.append(jrow_coh)
print(
datetime.datetime.now(),
'[%s/%s] %s %s %s - mean: %.3f, max: %.3f' %
(i2e, nbr_experiments, dataset, black_box, k,
jrow_coh['mean'], jrow_coh['max']))
except Exception:
print('error instance to explain: %d' % i2e)
errors.write('%d\n' % i2e)
continue
results = open(results_filename, 'a')
for jrow in jrow_list:
results.write('%s\n' % json.dumps(jrow))
results.close()
def main():
random_state = 0
dataset = 'fashion'
black_box = 'RF'
ae_name = 'aae'
path = './'
path_models = path + 'models/'
path_aemodels = path + 'aemodels/%s/%s/' % (dataset, ae_name)
black_box_filename = path_models + '%s_%s' % (dataset, black_box)
_, _, X_test, Y_test, use_rgb = get_dataset(dataset)
bb_predict, bb_predict_proba = get_black_box(black_box, black_box_filename,
use_rgb)
ae = get_autoencoder(X_test, ae_name, dataset, path_aemodels)
ae.load_model()
class_name = 'class'
class_values = ['%s' % i for i in range(len(np.unique(Y_test)))]
i2e = 10
img = X_test[i2e]
explainer = ILOREM(bb_predict,
class_name,
class_values,
neigh_type='rnd',
use_prob=True,
size=1000,
ocr=0.1,
kernel_width=None,
kernel=None,
autoencoder=ae,
use_rgb=use_rgb,
valid_thr=0.5,
filter_crules=True,
random_state=random_state,
verbose=True,
alpha1=0.5,
alpha2=0.5,
metric=neuclidean,
ngen=10,
mutpb=0.2,
cxpb=0.5,
tournsize=3,
halloffame_ratio=0.1,
bb_predict_proba=bb_predict_proba)
exp = explainer.explain_instance(img,
num_samples=1000,
use_weights=True,
metric=neuclidean)
print('e = {\n\tr = %s\n\tc = %s \n}' % (exp.rstr(), exp.cstr()))
print(exp.bb_pred, exp.dt_pred, exp.fidelity)
print(exp.limg)
img2show, mask = exp.get_image_rule(features=None, samples=10)
if use_rgb:
plt.imshow(img2show, cmap='gray')
else:
plt.imshow(img2show)
bbo = bb_predict(np.array([img2show]))[0]
plt.title('image to explain - black box %s' % bbo)
plt.show()
# if use_rgb:
# plt.imshow(img2show, cmap='gray')
# else:
# plt.imshow(img2show)
dx, dy = 0.05, 0.05
xx = np.arange(0.0, img2show.shape[1], dx)
yy = np.arange(0.0, img2show.shape[0], dy)
xmin, xmax, ymin, ymax = np.amin(xx), np.amax(xx), np.amin(yy), np.amax(yy)
extent = xmin, xmax, ymin, ymax
cmap_xi = plt.get_cmap('Greys_r')
cmap_xi.set_bad(alpha=0)
# Compute edges (to overlay to heatmaps later)
percentile = 100
dilation = 3.0
alpha = 0.8
xi_greyscale = img2show if len(img2show.shape) == 2 else np.mean(img2show,
axis=-1)
in_image_upscaled = transform.rescale(xi_greyscale,
dilation,
mode='constant')
edges = feature.canny(in_image_upscaled).astype(float)
edges[edges < 0.5] = np.nan
edges[:5, :] = np.nan
edges[-5:, :] = np.nan
edges[:, :5] = np.nan
edges[:, -5:] = np.nan
overlay = edges
# abs_max = np.percentile(np.abs(data), percentile)
# abs_min = abs_max
# plt.pcolormesh(range(mask.shape[0]), range(mask.shape[1]), mask, cmap=plt.cm.BrBG, alpha=1, vmin=0, vmax=255)
plt.imshow(mask,
extent=extent,
cmap=plt.cm.BrBG,
alpha=1,
vmin=0,
vmax=255)
plt.imshow(overlay,
extent=extent,
interpolation='none',
cmap=cmap_xi,
alpha=alpha)
plt.axis('off')
plt.title('attention area respecting latent rule')
plt.show()
def main():
dataset = sys.argv[1]
black_box = sys.argv[2]
neigh_type = sys.argv[3]
if len(sys.argv) > 4:
start_from = int(sys.argv[4])
else:
start_from = 0
# dataset = 'mnist'
# black_box = 'DNN'
# neigh_type = 'hrgp'
random_state = 0
ae_name = 'aae'
nbr_experiments = 200
if dataset not in ['mnist', 'cifar10', 'fashion']:
print('unknown dataset %s' % dataset)
return -1
if black_box not in ['RF', 'AB', 'DNN']:
print('unknown black box %s' % black_box)
return -1
if neigh_type not in ['rnd', 'gntp', 'hrgp']:
print('unknown neigh type %s' % neigh_type)
return -1
path = './'
path_models = path + 'models/'
path_results = path + 'results/fcp/'
path_aemodels = path + 'aemodels/%s/%s/' % (dataset, ae_name)
path_neigh = './neigh/'
black_box_filename = path_models + '%s_%s' % (dataset, black_box)
results_filename = path_results + 'alore_p_%s_%s_%s.json' % (
dataset, black_box, neigh_type)
neigh_filename = path_neigh + 'alore_%s_%s_%s.json.gz' % (
dataset, black_box, neigh_type)
_, _, X_test, Y_test, use_rgb = get_dataset(dataset)
bb, transform = get_black_box(black_box,
black_box_filename,
use_rgb,
return_model=True)
bb_predict, bb_predict_proba = get_black_box(black_box, black_box_filename,
use_rgb)
ae = get_autoencoder(X_test, ae_name, dataset, path_aemodels)
ae.load_model()
class_name = 'class'
class_values = ['%s' % i for i in range(len(np.unique(Y_test)))]
explainer = ILOREM(bb_predict,
class_name,
class_values,
neigh_type=neigh_type,
use_prob=True,
size=1000,
ocr=0.1,
kernel_width=None,
kernel=None,
autoencoder=ae,
use_rgb=use_rgb,
valid_thr=0.5,
filter_crules=True,
random_state=random_state,
verbose=False,
alpha1=0.5,
alpha2=0.5,
metric=neuclidean,
ngen=10,
mutpb=0.2,
cxpb=0.5,
tournsize=3,
halloffame_ratio=0.1,
bb_predict_proba=bb_predict_proba)
errors = open(
path_results + 'errors_alore_%s_%s_%s.csv' %
(dataset, black_box, neigh_type), 'w')
for i2e in range(nbr_experiments):
if i2e < start_from:
continue
img = X_test[i2e]
start_time = datetime.datetime.now()
try:
exp = explainer.explain_instance(img,
num_samples=1000,
use_weights=True,
metric=neuclidean)
except Exception:
print('error instance to explain: %d' % i2e)
errors.write('%d\n' % i2e)
errors.flush()
continue
run_time = (datetime.datetime.now() - start_time).total_seconds()
fidelity = exp.fidelity
img_cdist = transform(np.array([img]))
Zl = exp.Z
Z = exp.autoencoder.decode(Zl)
Z_cdist = transform(Z)
if black_box == 'DNN':
img_cdist = np.array([x.ravel() for x in img_cdist])
Z_cdist = np.array([x.ravel() for x in Z_cdist])
rdist = cdist(img_cdist, Z_cdist, metric='euclidean')
compact, compact_var = float(np.mean(rdist)), float(np.std(rdist))
sdist = cdist(np.array([exp.Z[0]]), exp.Z, metric=neuclidean)
lcompact, lcompact_var = float(np.mean(sdist)), float(np.std(sdist))
X_test_cdist = transform(X_test)
if black_box == 'DNN':
X_test_cdist = np.array([x.ravel() for x in X_test_cdist])
dist = cdist(img_cdist, X_test_cdist, metric='euclidean')
nbr_real_instances = len(X_test)
plausibility = calculate_plausibilities(rdist, dist,
nbr_real_instances)
print(
datetime.datetime.now(),
'[%s/%s] %s %s %s - f: %.2f, c: %.2f, lc: %.2f, p: %.2f' %
(i2e, nbr_experiments, dataset, black_box, neigh_type, fidelity,
compact, lcompact, plausibility[-2]))
store_fcpn(i2e, results_filename, neigh_filename, dataset, black_box,
fidelity, compact, compact_var, lcompact, lcompact_var,
plausibility, run_time, Z, Zl, neigh_type)
errors.close()