some inconsistency with the paper

[Okazaki86]

1. in the paper it calculates class centers within a single image, but implements as within a batch, which just return a matrix of (class, feature dim) for method dec_deeplabv3_contrast.contrast_region.
2. the variable l_pos in dec_deeplabv3_contrast.forward when calculating contrast loss should be a similarity score of 1 dim, but it shows still 256 dim in the code.

[liangfengyouxingii]

@Okazaki86 hello, I also notice this problem, do you solve it?

[liangfengyouxingii]

may l_pos = query.unsqueeze(0) @ eval("self.queue" + str(cls_ind)).clone().detach()

[Okazaki86]

@Okazaki86 hello, I also notice this problem, do you solve it?

我按论文里的方法进行了改正