def calculate_score(loss_file, reverse, smoothing):
if not os.path.isdir(loss_file):
loss_file_path = loss_file
else:
optical_result = compute_auc(loss_file, reverse, smoothing)
loss_file_path = optical_result.loss_file
print('##### optimal result and model = {}'.format(optical_result))
dataset, psnr_records, gt = load_psnr_gt(loss_file=loss_file_path)
# the number of videos
num_videos = len(psnr_records)
scores = np.array([], dtype=np.float32)
labels = np.array([], dtype=np.int8)
# video normalization
for i in range(num_videos):
distance = psnr_records[i]
distance = (distance - distance.min()) / (distance.max() -
distance.min())
if reverse:
distance = 1 - distance
if smoothing:
distance = score_smoothing(distance)
scores = np.concatenate(
(scores[:], distance[DECIDABLE_IDX:-DECIDABLE_IDX]), axis=0)
labels = np.concatenate(
(labels[:], gt[i][DECIDABLE_IDX:-DECIDABLE_IDX]), axis=0)
mean_normal_scores = np.mean(scores[labels == 0])
mean_abnormal_scores = np.mean(scores[labels == 1])
print('mean normal scores = {}, mean abnormal scores = {}, '
'delta = {}'.format(mean_normal_scores, mean_abnormal_scores,
mean_normal_scores - mean_abnormal_scores))
def get_scores_labels(loss_file, reverse, smoothing):
# the name of dataset, loss, and ground truth
dataset, psnr_records, gt = load_psnr_gt(loss_file=loss_file)
# the number of videos
num_videos = len(psnr_records)
scores = np.array([], dtype=np.float32)
labels = np.array([], dtype=np.int8)
# video normalization
for i in range(num_videos):
distance = psnr_records[i]
if NORMALIZE:
distance = (distance - distance.min()) / (distance.max() -
distance.min() + 1e-8)
#distance -= distance.min() # distances = (distance - min) / (max - min)
#distance /= distance.max()
if reverse:
distance = 1 - distance
if smoothing:
distance = score_smoothing(distance)
scores = np.concatenate(
(scores[:], distance[DECIDABLE_IDX:-DECIDABLE_IDX]), axis=0)
labels = np.concatenate(
(labels[:], gt[i][DECIDABLE_IDX:-DECIDABLE_IDX]), axis=0)
return dataset, scores, labels
def compute_auc_average(loss_file, reverse, smoothing):
if not os.path.isdir(loss_file):
loss_file_list = [loss_file]
else:
loss_file_list = os.listdir(loss_file)
loss_file_list = [
os.path.join(loss_file, sub_loss_file)
for sub_loss_file in loss_file_list
]
optimal_results = RecordResult()
for sub_loss_file in loss_file_list:
# the name of dataset, loss, and ground truth
dataset, psnr_records, gt = load_psnr_gt(loss_file=sub_loss_file)
if dataset == 'shanghaitech':
gt[51][5] = 0
elif dataset == 'ped2':
for i in range(7, 11):
gt[i][0] = 0
elif dataset == 'ped1':
gt[13][0] = 0
# the number of videos
num_videos = len(psnr_records)
scores = np.array([], dtype=np.float32)
labels = np.array([], dtype=np.int8)
# video normalization
auc = 0
for i in range(num_videos):
distance = psnr_records[i]
if NORMALIZE:
distance = (distance - distance.min()) / (
distance.max() - distance.min() + 1e-8)
# distance -= distance.min() # distances = (distance - min) / (max - min)
# distance /= distance.max()
# for the score is anomaly score
if reverse:
distance = 1 - distance
# to smooth the score
if smoothing:
distance = score_smoothing(distance)
# scores = np.concatenate((scores[:], distance[DECIDABLE_IDX:-DECIDABLE_IDX]), axis=0)
# labels = np.concatenate((labels[:], gt[i][DECIDABLE_IDX:-DECIDABLE_IDX]), axis=0)
#_auc = metrics.roc_auc_score(np.array(gt[i],dtype=np.int8),np.array(distance,dtype=np.float32))
#_auc = metrics.auc(fpr, tpr)
fpr, tpr, thresholds = metrics.roc_curve(np.array(gt[i],
dtype=np.int8),
np.array(
distance,
dtype=np.float32),
pos_label=0)
_auc = metrics.auc(fpr, tpr)
print('video {}: auc is {}'.format(i + 1, _auc))
auc += _auc
auc /= num_videos
print(auc)
def compute_auc(loss_file, reverse, smoothing):
if not os.path.isdir(loss_file):
loss_file_list = [loss_file]
else:
loss_file_list = os.listdir(loss_file)
loss_file_list = [
os.path.join(loss_file, sub_loss_file)
for sub_loss_file in loss_file_list
]
optimal_results = RecordResult()
for sub_loss_file in loss_file_list:
# the name of dataset, loss, and ground truth
dataset, psnr_records, gt = load_psnr_gt(loss_file=sub_loss_file)
# the number of videos
num_videos = len(psnr_records)
scores = np.array([], dtype=np.float32)
labels = np.array([], dtype=np.int8)
# video normalization
for i in range(num_videos):
distance = psnr_records[i]
if NORMALIZE:
distance = (distance - distance.min()) / (
distance.max() - distance.min() + 1e-8)
# distance -= distance.min() # distances = (distance - min) / (max - min)
# distance /= distance.max()
# for the score is anomaly score
if reverse:
distance = 1 - distance
# to smooth the score
if smoothing:
distance = score_smoothing(distance)
scores = np.concatenate(
(scores[:], distance[DECIDABLE_IDX:-DECIDABLE_IDX]), axis=0)
labels = np.concatenate(
(labels[:], gt[i][DECIDABLE_IDX:-DECIDABLE_IDX]), axis=0)
fpr, tpr, thresholds = metrics.roc_curve(labels, scores, pos_label=0)
auc = metrics.auc(fpr, tpr)
results = RecordResult(fpr, tpr, auc, dataset, sub_loss_file)
if optimal_results < results:
optimal_results = results
if os.path.isdir(loss_file):
print(results)
print('##### optimal result and model AUC= {}'.format(optimal_results))
return optimal_results