About args n_view,max_samples

[xianxuan-z]

Hello! Your work is outstanding! I did a detailed research on both the paper and the code. There are a lot of questions that I don’t understand:

1. I want to ask first about the loss_contrast.py Can you explain the meaning of max_samples and n_views in the hard_anchor_sampling function?
2. If you apply this method to small samples, can the hard_anchor_sampling step be omitted? I hope you can explain in more detail，many thanks！

[tfzhou]

Hi，Thanks for your interest.

1. As given in the following line, max_samples is the maximum number of samples for all classes, and n_views is the average number for each class. https://github.com/tfzhou/ContrastiveSeg/blob/7be326ae9127a847558584eac489d4858adf7cb1/lib/loss/loss_contrast.py#L47
2. You can also try to find some hard anchors (which are misclassified), no matter how many samples you have.

[xianxuan-z]

Thank you for your prompt reply! Then I have corresponding questions and look forward to your help to explain: classes = [] total_classes = 0 for ii in range(batch_size): this_y = y_hat[ii] this_classes = torch.unique(this_y) this_classes = [x for x in this_classes if x != self.ignore_label] this_classes = [x for x in this_classes if (this_y == x).nonzero().shape[0] > self.max_views]

        classes.append(this_classes)
        total_classes += len(this_classes)

In your code, if there are the same categories in different batches, will there be duplicate statistics in the result of total_classes? Hope you can solve my confusion.

[tfzhou]

What do you mean for "duplicate statistics"? p.s. If you have many questions, please also feel free to compile them together here at one time.

[zh-aidy]

Hello! Your work is outstanding! I did a detailed research on both the paper and the code. There are a lot of questions that I don’t understand:

1. Can you explain the meaning of self.max_views in the hard_anchor_sampling function？ And why need (this_y == x).nonzero().shape[0] > self.max_views ？
2. In my project, some labels only have a little targat pixels, specially resized to the size of embedding, so the total_classes==0 situation may be happened, so that feats_ is None and self._contrast function can not run.
3. I want to ask about the class HRNet_W48_MEM of hrnet.py: def _momentum_update_key_encoder(self): for param_q, param_k in zip(self.encoder_q.parameters(), self.encoder_k.parameters()): param_k.data = param_k.data * self.m + param_q.data * (1. - self.m) Can you explain the meaning of self.encoder_k？I am sorry that I did not find its definition. I hope you can explain in more detail，many thanks！