Wenbin Li
Wenbin Li
Thanks. No, DN4 does not need a pre-training stage and it is just trained from scratch in this code. By the way, since pre-training is a general trick, you can...
You are welcome. As mentioned in our paper, one key reason is that we employ the much richer non-summarized local representations to represent both the query image and support class....
It's my pleasure. I am glad that DN4 works on your dataset! Yes, Transformer is a good choice. Unfortunately, I don't have much good suggestions or experiences on this part....
Yes, it's a normal situation. Because DN4 use a Sum operation to aggregate all the local similarities for a query image, this will make the similarity list flat. Fortunately, the...
Sorry for the late reply. Do you still need the implementation of prototypical network?
您好,关于第一个问题,是的,在训练阶段进行了测试,方便观察模型的效果; 关于第二个问题,这样设置的目的是希望模型训练完一个Epoch之后,固定模型BN层的参数,提高模型的泛化能力,是一个小的trick,当然你可以去掉这个设置,但是需要在train代码里把模型的模式改成model.train()。
EpisodeSize代表训练时few-shot task的个数,从task维度理解,batch_size=1,但是如果从sample维度理解,batch_size=support samples + query samples; 改变它的大小对训练有一定的影响,看你是增加task的个数,还是说只是改变query samples的个数,一般不会改变support set的大小。 谢谢! --------------------------------------------------- Wenbin Li, PhD. Assistant Professor Department of Computer Science and Technology State Key Laboratory for Novel Software Technology Nanjing...
您好,在dog和bird数据集上是否存在严重的过拟合现象呢? 验证集结果和测试集结果的差异可能是固定BN造成的,您看看是否可以把CovaMNet_Train_5way1shot.py 里386行的model.eval() 换成model.train(),看看新的结果如何。 另外,由于dog和bird数据比较少,可能比较容易过拟合,其实应该做一下数据增广。
您好,这个问题问的非常好,我们其实是采用了一个比较巧妙的方式来实现这个可学习权重w, 我们在计算查询图像Q与某个类别的相似度的时候,假设Q有m个局部相似度,我们直接把这m个相似度存下来,放在mea_sim里,它的维度是"类别数量*m",然后在self.classifier里我们采用了一个Conv1d来,使得Kernel size和stride都等于m,即论文中的441,通过这种方式就自动学习了w nn.Conv1d(1, 1, kernel_size=441, stride=441, bias=use_bias)
您好,Table 2里是完全按照原文的代码里的设置和使用各种trick,但是Table 3是尽可能让所有方法使用差不多的设置和相同的训练trick,包括训练的轮数等,主要是这个原因造成的,这样可能也不是非常公平公正,但已经是相对合理的了。