def make_MCpredictionBIS(self, input_img, args):
im_shape = np.shape(input_img)
if len(input_img.shape) != 4:
input_img = np.array([input_img])
args['4_MCdrop'] = True
my_dataset, new_net = instantiate_net(args, Train=False)
new_net.MCDropout = True
layer_dict = dict([(layer.name, layer) for layer in self.model.layers])
newlayers = [layer.name for layer in new_net.model.layers]
for idx, layer_name in enumerate(newlayers):
if 'Dropout' not in layer_name:
if layer_name in layer_dict and len(
layer_dict[layer_name].get_weights()) > 0:
new_net.model.layers[idx].set_weights(
layer_dict[layer_name].get_weights())
n_pred = 30
mean_pred, var_pred = np.zeros(
input_img.shape, dtype='uint8'), np.zeros(input_img.shape)
for i, input_im in enumerate(input_img):
these_pred = np.zeros(
(n_pred, input_im.shape[0], input_im.shape[1],
1 if self.Grayscale else 3),
dtype='uint8')
for j in range(n_pred):
these_pred[j] = new_net.make_prediction(input_im)
mean_pred[i] = np.mean(these_pred, axis=0)
var_pred[i] = np.var(these_pred, axis=0)
return mean_pred, var_pred
def call_training(this_exp):
my_dataset, my_net = instantiate_net(this_exp)
print(
colored('Start to train the network for experiment : %s.', 'red') %
(my_net.exp_name))
my_net.fit(this_exp, my_dataset)
print(colored('The network is trained.', 'red'))
my_net.print_learning_curves()
print('curves printed')
my_net.load_weights(my_net.exp_name)
my_net.evaluate(my_dataset)
def evaluate_net(args):
if args['4_ROC'] or args['4_ROC_pixel']:
for this_exp in args['4_exp']:
this_args = read_json(root_path + '/Experiments/' + this_exp +
'/00_description.json')
my_dataset, my_net = instantiate_net(this_args)
my_net.load_weights(this_exp)
my_net.evaluate(my_dataset,
print_pred=args['4_ROC_printpred'],
image_wise=args['4_ROC'],
pixel_wise=args['4_ROC_pixel'])
if args['4_MCdrop'] or args['4_MCdrop_pixel']:
for this_exp in args['4_exp']:
this_args = read_json(root_path + '/Experiments/' + this_exp +
'/00_description.json')
this_args['4_MCdrop'] = True
_, my_net = instantiate_net(this_args)
my_net.load_weights(this_exp)
my_net.evaluate_MCDropout(this_args,
print_pred=args['4_MCdrop_printpred'],
image_wise=args['4_MCdrop'],
pixel_wise=args['4_MCdrop_pixel'])
def evaluate_MCDropout(self, args, print_pred=True, pixel_wise = False, ROC=True, norms=['l0', 'l1', 'l2', 'linf']):
args['4_MCdrop'] = True
my_dataset, new_net = instantiate_net(args, Train=False)
new_net.MCDropout = True
# Compute the predictions
used = set()
subsets = [x for x in my_dataset.test_label_name if x not in used and (used.add(x) or True)]
all_mask = {}
for this_subset in subsets:
all_mask[this_subset] = my_dataset.mask[my_dataset.test_label_name==this_subset]
# Infer all the test images
print('-------start Inference---------')
print('xxxxxx', self.exp_name)
all_pred = self.make_MCprediction(my_dataset, new_net, print_pred)
print('-------end Inference---------')
# Compute the ROC curve for each type of anomaly
subsets.remove('clean')
for this_subset in subsets:
y = np.concatenate( (all_pred['clean']['variance'], all_pred[this_subset]['variance']) )
x = np.zeros(y.shape)
masks = np.concatenate( (all_mask['clean'], all_mask[this_subset]) )
binary_test_labels = my_dataset.test_label
binary_test_labels = np.concatenate(( my_dataset.test_label[my_dataset.test_label_name=='clean'], my_dataset.test_label[my_dataset.test_label_name==this_subset]))
binary_test_labels[binary_test_labels ==2] = 1
if not pixel_wise:
all_TP, all_FP = compute_ROC(x, y, binary_test_labels)
result_path = root_path + '/Results_MCDropout/' + self.exp_name + '/ROC_clean_vs_' + this_subset
write_json({'all_TP': all_TP, 'all_FP': all_FP}, result_path + '.json')
print_ROC(result_path + '.json', result_path + '.png', AUC=True)
else :
print('------', this_subset)
TP, FP = compute_ROC_pixel_wise(x, y, masks)
result_path = root_path + '/Results_MCDropout/' + self.exp_name + '/PIXELWISE_ROC_clean_vs_' + this_subset
write_json({'TP': TP, 'FP': FP}, result_path + '.json')
print_PIXEL_WISE_ROC(result_path + '.json', result_path + '.png', AUC=True)
# Compute combined ROC curve
all_def = my_dataset.test_label_name[my_dataset.test_label ==2]
used = set()
subsets = [x for x in all_def if x not in used and (used.add(x) or True)]
prediction = None
for this_subset in subsets :
if prediction is None:
prediction = all_pred[this_subset]['variance']
else:
prediction = np.vstack( (prediction, all_pred[this_subset]['variance']))
y = np.concatenate( ( all_pred['clean']['variance'], prediction) )
x = np.zeros(y.shape)
masks = np.concatenate((my_dataset.mask[my_dataset.test_label_name=='clean'], my_dataset.mask[my_dataset.test_label==2]))
binary_test_labels = np.concatenate(( my_dataset.test_label[my_dataset.test_label_name=='clean'], my_dataset.test_label[my_dataset.test_label==2]))
binary_test_labels[binary_test_labels ==2] = 1
if not pixel_wise:
all_TP, all_FP = compute_ROC(x, y, binary_test_labels)
result_path = root_path + '/Results_MCDropout/' + self.exp_name + '/ROC_clean_vs_realDefaults'
write_json({'all_TP': all_TP, 'all_FP': all_FP}, result_path + '.json')
print_ROC(result_path + '.json', result_path + '.png', AUC=True)
else :
TP, FP = compute_ROC_pixel_wise(x, y, masks)
result_path = root_path + '/Results_MCDropout/' + self.exp_name + '/PIXELWISE_ROC_clean_vs_realDefaults'
write_json({'TP': TP, 'FP': FP}, result_path + '.json')
print_PIXEL_WISE_ROC(result_path + '.json', result_path + '.png', AUC=True)