Yutong Lin

Hi, the shared checkpoints are NOT tar files, but pre-trained weights. Sorry for the misleading postfix. We can load these pre-trained weights in this way: 

HI, @gkgo . Thanks for your interest in this work. Could you share the command you used with us to reproduce the error?

> I really appreciate your reply. The problem I encountered is：  Do you use a command like this: ``` python main.py --config [CONFIGFILENAME] \ --supp [SUPPLEMENTSTRING] \...

We have not tested our code under Windows and you could have a try. In addition, you need to replace "[CONFIGFILENAME]" "[SUPPLEMENTSTRING]" ... with specific values. [This folder](https://github.com/bl0/negative-margin.few-shot/tree/master/script) provides some...

Hi, do you mean the value of loss is negative? If this is the point, this is right for PixPro. You may check the implementation of this line https://github.com/zdaxie/PixPro/blob/main/contrast/models/PixPro.py#L204 .

Hi, @UcanSee . We set the same distance threshold for every level of FPN.

Hi, @Alxead . From our experience, a "sqrt" scheduling method should be used to adjust the learning rate. As our default setting, the learning rate for batch size 1024 is:...

Hi, @aaronsarna . We use the 8th layer of ViT for linear evaluation. In addition, we do not use the [CLS] token. Instead, we apply average pooling over the image...