The number of label without anno

[DuanchuWang]

`       
if self.split=='train' or self.split=='trainval':
#print('----------------------')
if '%' in self.labeled_point:
    r = float(self.labeled_point[:-1]) / 100
    num_pts = len(data[0]) #data.shape[0]
    num_with_anno = max(int(num_pts * r), 1)
    num_without_anno = num_pts - num_with_anno
    idx_without_anno = np.random.choice(num_pts, num_without_anno, replace=False)
    label[idx_without_anno]= -100 #Unlabeled
else:
    num_pts = len(data[0]) #data.shape[0]
    num_with_anno = int(self.labeled_point)
    num_without_anno = num_pts - num_with_anno
    idx_without_anno = np.random.choice(num_pts, num_without_anno, replace=False)
    label[idx_without_anno] = -100

I think only the percentage of weakly supervised labels is fixed here, but due to random sampling, the weakly supervised labels are not actually fixed. Is this correct?

[LinLin1031]

弱监督学习中的弱标签一般都是按照百分比随机选的吧。

[DuanchuWang]

弱监督学习中的弱标签一般都是按照百分比随机选的吧。

这些code出现在__getitem__里，每次调用都会重新选择真的没问题吗？是不是应该固定下来

[LinLin1031]

弱监督学习中的弱标签一般都是按照百分比随机选的吧。

这些code出现在__getitem__里，每次调用都会重新选择真的没问题吗？是不是应该固定下来

您这么说也有道理，弱监督方法使用的标签是有限的标注，这些标注在训练中应该是不可能改变的。