negtive sample choose problem

[xuchengggg]

In def ssd_losses_old, nvalues = tf.where(nmask,predictions[:, :, :, :, 0],1. - fnmask), so nvalues = 1 if this anchor is positive and navalues = predictions[:,:,:,:,0], and is this predictions indicate the possibilities of background, and then val, idxes = tf.nn.top_k(-nvalues_flat, k=n_neg), is this meanings the negative samples are chosen which have low possibilities of background first?

[understand94]

首先nvalues的意思就搞错了。nvalues在positive的时候为0，反之，对应为背景类的概率。top_k本身是取前K大的数，现在nvalues取反了，也就是说，idxes里面是最小的K个数，idxes[-1]指的是这K个小的数中最大的，在后面的代码直接用这个最大的数来筛选。