search

Fangyi-Chen / SQR

MIT License
102 stars 5 forks source link

Deformable DETR code #2

Closed yuzitian closed 1 year ago

yuzitian commented 1 year ago

Hi, thank you for a good job. You did experiments with DETR-like models in your paper. I would like to ask if it is convenient for you to open the source code of the experiment with Deformable DETR?

Fangyi-Chen commented 1 year ago

Hi,

Yes we plan to release sqr-deformable detr in a couple of weeks as the team is on leave

Fangyi-Chen commented 1 year ago

[update] Released. Please refer to the readme.md

Recommend following the original deformable detr setting (32 as batchsize 8*A100 x 4 imgs per GPU ), but it seems 16 as batchsize performs similarly. Expecting 45.8±0.2 AP. Thanks

yuzitian commented 1 year ago

Thank you very much for your reply. I ran your code as soon as you replied. Due to insufficient memory on my device, I am unable to use the same configuration as you. But I got a bad result. After carefully reviewing the code, I found that you did not update the reference points during the decoder process. May I ask why you are handling this in this place?

Fangyi-Chen commented 1 year ago

Hi,

I am not sure what hyperparameter you changed in config, but some of them are critical, such as query number, feature dimension, encoder layer number, decoder layer number, etc.

For sqr-deformable-detr, we do not collect the reference points, because if the reg_branch is None (one stage), the reference points are not updated, so for SQR, it is also unnecessary to do that.

If you do not have enough memory to run your research, I recommend using sqr-adamixer which runs well with 8x24GB card.

yuzitian commented 1 year ago

Thank you a lot for your advice. Wish you have a nice day !!