Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Here is a bug? #26

Open
August-en opened this issue Nov 16, 2022 · 5 comments
Open

Here is a bug? #26

August-en opened this issue Nov 16, 2022 · 5 comments

Comments

@August-en
Copy link

cur_pos_embed = lvl_pos_embed_flatten[0:1]
ref_pos_embed_list = torch.chunk(lvl_pos_embed_flatten[1:], self.num_ref_frames, dim=0)
ref_pos_embed = torch.cat(ref_pos_embed_list, 1)
ref_memory = ref_memory + ref_pos_embed

@August-en
Copy link
Author

August-en commented Nov 16, 2022

This will be all right?

ref_pos_embed_list = torch.chunk(lvl_pos_embed_flatten, self.num_ref_frames+1, dim=0)
cur_pos_embed = lvl_pos_embed_flatten[0]
ref_pos_embed = torch.cat(ref_pos_embed_list[1:], 1)
ref_memory = ref_memory + ref_pos_embed

@akanuasiegbu
Copy link

I get this error as well when the batch size is greater than one. When batch size is one this error does not appear. The code on the repo does not appear to work for an arbitrary batch size currently

@Zagreus98
Copy link

Try with this PR #13 for batchsize > 1. I tested it and it works, though would not recommend if you do not have gpus with > 32GB memory.

@August-en
Copy link
Author

Try with this PR #13 for batchsize > 1. I tested it and it works, though would not recommend if you do not have gpus with > 32GB memory.

Thank you so much. I will try it as soon as possible.

Have you used the TDTE module in your expriments? I found that the default setting about TDTD in this repo is False. (#27 (comment)). Does it make a big difference whether to use it or not?

Thank you again if you can share your experience :)

@Zagreus98
Copy link

Zagreus98 commented Nov 18, 2022

Yes, I did experiments with and without TDTE on my own dataset and the performance was almost the same. What worked in my case to increase performance was to add Illumination variation augmentation and class weights in the loss. I also reproduced the results on ImageNet VID from this page with TDTD set to False, therefore I don't know if it worth adding it or not, it may depend on your dataset.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

3 participants