AvadaKarrot

Results 5 comments of AvadaKarrot

I have the same questions as well. In general criteria, when in inference phase as far as I know, it is not allowed to get access to Test Data Label...

> 这个应该是类似于multi query的 map,因为使用了centroids其实就是相当于利用了多个gallery的信息去做召回,不一样的是,一个是multi query,另外一个是multi gallery。不知道这样理解对不对。 我觉着这个理解是对的,reranking的方法改变了. 但是一般现实情况下,只能拿到一张query,而不是多张query去在Gallery里寻找. 这个方法排序出来的结果肯定是会更好的, 避免了一些hard样本排序的结果.

Have you successfully figured out these problems? I encountered them either. Could you please tell me how to set the hyper-parameters. Thank u.

您好,请问您解决这个问题了吗? 我发现visualizer可视化的是某一个transformer layer的某个attention head的attn map,attn_map[i]的shape是(1, 12, 129, 129), grid_index是从0-128的int,在visualizer.py这个函数里,attention_map是(129, )的向量,如果grid_index取0,是不是就是cls的特征图,可以理解成全局的注意力图吗? 我也想要全局注意力图,但是没想出合理的办法。请不吝赐教~

I suppose it is okay to interpolate (torch.nn.functional.interpolate) the complex_weight. However, in my task, changing it seems affect the model capacity.