DavideHe

> We do not replace the first and last conv in the networks, since replacing them would affect the performance and their computational cost is relatively very small in the...

> Hi, > > Thank you for your comment. However, I'm not really getting your point. Are you suggesting to introduce a different loss function? Are you experiencing problems with...

> Thanks for your feedback. I can see the reason for your concern. However, my implementation aimed to be faithful to the original implementation in the paper, which did not...

> **描述该问题** from baiduspider import BaiduSpider 出现如下： ImportError: cannot import name 'BaiduSpider' from partially initialized module 'baiduspider' (most likely due to a circular import) > > **预期结果** > > 能正常运行...

> “先采用transforms.Resize(256)，然后transforms.CenterCrop(224)”这种方式是普遍采用的方式。这种方式是保持长宽比的情况下将短边resize到256，然后中心crop出224x224的图片，每次产生的图片是一样的。没有你说的不一样的情况。 抱歉，我将CenterCrop理解成RandomResizedCrop了； 但是实际工程中，一般是不会crop吧，crop对分类的主体在中心位置是友好的，但是主体不在中心位置并不好，并且对于长宽比大幅偏离1：1，这样的操作可能直接丢失主体。

感觉可以所有的卷积(包括1x1的)都换成adder，不要一会addernet一会一般卷积，因为addernet是计算1范式距离，但是1x1的卷积算的是余弦距离，混用可能导致不好的结果