When I train my code, it did not work. How to sove the problem, please?Always stuck in this interface
My environment:

In the 'softmax_with_policy' function:
attn_policy = attn_policy + (1.0 - attn_policy) * eye
Is the second item ( (1.0 - attn_policy) * eye) equal to 0

For stable training

    max_att, _ = torch.max(attn, dim=-1, keepdim=True)
    attn = attn - max_att
    attn = attn.to(torch.float32).exp_() * attn_policy.to(torch.float32)
    attn = (attn + eps / N) / (attn.sum(dim=-1, keepdim=True) + eps)

hello ! i want to know how you test speed on your model (45fps on got10k)? can u provide the "profile_model.py" like other tracker? thank u!

hi, congratulations! great work.
In your papaer ,I'm a little confused about formula 7

If the Di is tempalte token ,it should be (1,0,0), and weather the Dj is ES token or EA token , by the 7th formula ,the Mij is zero.

So how do ET and EA interact, refer to the second formula

I am looking forward to your reply!Thank you.

Could you please explain why certain datasets require a threshold and why there are different thresholds for them during inference?
` if self.training:
# During training
decision = F.gumbel_softmax(divide_prediction, hard=True)
else:
# During inference
if threshold:
# Manual rank based selection
decision_rank = (F.softmax(divide_prediction, dim=-1)[:, :, 0] < threshold).long()
else:
# Auto rank based selection
decision_rank = torch.argsort(divide_prediction, dim=-1, descending=True)[:, :, 0]

            decision = F.one_hot(decision_rank, num_classes=2)`

用的单个GPU，训练的结果比你这差了五六个点，是为什么？有没有可以改进的办法。

Hi, thank you for your great work.
Is there any inference code or demo to run your work on a custom video?
can I implement it by using lib.test.evaluation.tracker.py ?

你好，请问GRM-L320单独在got10k数据集上训练和测试的效果怎么样？

It seems like the batch size is 40 on each GPU, and the learning rate is 0.0004 in total in the provided config. How should I set the learning rate and the batch size if I have only one RTX3090 GPU? BTW, could u please share ur training log? I just wonder how to confirm when it is converged.

When I test the model on nfs30, the architecture of nfs30 downloaded from the official website is different from the requirement of the code, then I reorganize the structures of the dataset as anno and sequence, however, I meet the error: the label in the txt file can not convert to float.

My question is that should I reorganize the architecture of nfs30? or where can I download the NFS30 dataset with anno and sequence structures?

Hi, congratulations, this is a nice job! I have a question about Appendix C3. In the paper, it can use the continuous estimation to scale the raw attention weights that bypass the problem of the non-differentiable obstacle. But how to do that? Can you give me some references?
Hope your reply!

训练代码的时候出现“ValueError: The number of weights does not match the population”，更换numpy的版本也不行，请问您有遇到过这个问题吗？

Try training with trained_model -- GRM_ep0300.pth.tar provided by the author, but keep reporting error that checkpoint type is not the expected type (GRM)