def evaluate_cand_off(cand_with_off, ground_truth, ment_data, k=0.6):
recall = []
precision = []
f1 = []
under_k = []
error = []
error_off = []
for i in tqdm(range(len(ground_truth))):
tp = 0
for j in ground_truth[i]:
if j[0] in cand_with_off[i]:
if j[1] in cand_with_off[i][j[0]][0]:
tp += 1
else:
if j[1] in ment_data[i]:
error_off.append(i)
r = tp / len(ground_truth[i])
if r < k:
under_k.append(i)
p = tp / len(cand_with_off[i].keys())
f = safediv(2 * r * p, r + p)
recall.append(r)
precision.append(p)
if f > 1 or np.isnan(f):
error.append(i)
f1.append(0)
else:
f1.append(f)
av_recl = sum(recall) / len(recall)
av_pre = sum(precision) / len(precision)
av_f1 = sum(f1) / len(f1)
print('average recall: {}'.format(av_recl))
print('average precision: {}'.format(av_pre))
print('average f1: {}'.format(av_f1))
print('error number: {}'.format(len(error)))
print('total number: {}'.format(len(ment_data)))
print('error rate: {}'.format(len(error) / len(ment_data)))
print('error off: {}'.format(len(error_off)))
print('recall under{} : {}'.format(k, len(under_k)))
return recall, precision, f1, error, error_off, under_k
def tversky_upbound(aLen: int, aPassed: int, bLen: int, bPassed: int,
count: int) -> float:
return min(1.0,
safediv(overlap_upbound(**locals()), (1 - e) * aLen + e * bLen))
def jaccard_upbound(aLen: int, aPassed: int, bLen: int, bPassed: int,
count: int) -> float:
maxCount = overlap_upbound(**locals())
return min(1.0, safediv(maxCount, aLen + bLen - maxCount))
def tversky_upbound(aLen: int, aPassed: int, bLen: int, bPassed: int,
count: int) -> float:
# print('fenzi',overlap_upbound(**locals()))
# print('fenmu',(1 - e) * aLen + e * bLen)
return min(1.0,
safediv(overlap_upbound(**locals()), (1 - e) * aLen + e * bLen))