def extract_auprc(targets, scores, radius, threshold, match_radius=None, pool=None):
N = 0
mse = 0
hits = []
preds = []
if pool is not None:
process = ExtractMatches(radius, threshold, match_radius)
iterator = iterate_score_target_pairs(scores, targets)
for assignment,score,this_mse,n in pool.imap_unordered(process, iterator):
mse += this_mse
hits.append(assignment)
preds.append(score)
N += n
else:
for score,target in iterate_score_target_pairs(scores, targets):
score,coords = non_maximum_suppression(score, radius, threshold=threshold)
if match_radius is None:
assignment, dist = match_coordinates(target, coords, radius)
else:
assignment, dist = match_coordinates(target, coords, match_radius)
mse += np.sum(dist[assignment==1]**2)
hits.append(assignment)
preds.append(score)
N += len(target)
hits = np.concatenate(hits, 0)
preds = np.concatenate(preds, 0)
auprc = average_precision(hits, preds, N=N)
rmse = np.sqrt(mse/hits.sum())
return auprc, rmse, int(hits.sum()), N
def __call__(self, args):
score,target = args
score,coords = non_maximum_suppression(score, self.radius, threshold=self.threshold)
if self.match_radius is None:
assignment, dist = match_coordinates(target, coords, self.radius)
else:
assignment, dist = match_coordinates(target, coords, self.match_radius)
mse = np.sum(dist[assignment==1]**2)
return assignment, score, mse, len(target)
def main(args):
match_radius = args.assignment_radius
targets = pd.read_csv(args.targets, sep='\t')
predicts = pd.read_csv(args.predicted, sep='\t', comment='#')
if args.images == 'union':
image_list = set(targets.image_name.unique()) | set(predicts.image_name.unique())
elif args.images == 'target':
image_list = set(targets.image_name.unique())
elif args.images == 'predicted':
image_list = set(predicts.image_name.unique())
else:
raise Exception('Unknown image argument: ' + args.images)
image_list = list(image_list)
N = len(targets)
matches = []
scores = []
count = 0
mae = 0
for name in image_list:
target = targets.loc[targets.image_name == name]
predict = predicts.loc[predicts.image_name == name]
target_coords = target[['x_coord', 'y_coord']].values
predict_coords = predict[['x_coord', 'y_coord']].values
score = predict.score.values.astype(np.float32)
match,dist = match_coordinates(target_coords, predict_coords, match_radius)
this_mae = np.sum(dist[match==1])
count += np.sum(match)
delta = this_mae - np.sum(match)*mae
mae += delta/count
matches.append(match)
scores.append(score)
matches = np.concatenate(matches, 0)
scores = np.concatenate(scores, 0)
precision,recall,threshold,auprc = precision_recall_curve(matches, scores, N=N)
print('# auprc={}, mae={}'.format(auprc,np.sqrt(mae)))
mask = (precision + recall) == 0
f1 = 2*precision*recall
f1[mask] = 0
f1[~mask] /= (precision + recall)[~mask]
table = pd.DataFrame({'threshold': threshold})
table['precision'] = precision
table['recall'] = recall
table['f1'] = f1
table.to_csv(sys.stdout, sep='\t', index=False)
