search

IDEA-Research / detrex

detrex is a research platform for DETR-based object detection, segmentation, pose estimation and other visual recognition tasks.
https://detrex.readthedocs.io/en/latest/
Apache License 2.0
1.97k stars 206 forks source link

About training with DN-Deformable-DETR-R50 #120

Closed Mr-Bigworth closed 1 year ago

Mr-Bigworth commented 1 year ago

When I train DN-Deformable-DETR-R50 in 12 epochs, cause I only have one Tesla A-100 GPU, I set the dataloader.train.total_batch_size = 4 and train.max_iter = 360000. And the result of the AP and AP50 is 46.5559 and 64.1131 in iteration 334999 and the result seems to be better. Why it's so high... the detection accuracy.

FengLi-ust commented 1 year ago

Yes, it is possible. In our original DN-DETR repo, DN-Deformable-DETR can get 46.1 in 12 epochs.
One reason for your better performance is that training with small batchsize will accelerate convergence in the early stage, I have verified this before. In addition, detrex implementation is better than our original DN-DETR repo. (:

Mr-Bigworth commented 1 year ago

Yes, it is possible. In our original DN-DETR repo, DN-Deformable-DETR can get 46.1 in 12 epochs. One reason for your better performance is that training with small batchsize will accelerate convergence in the early stage, I have verified this before. In addition, detrex implementation is better than our original DN-DETR repo. (:

Thanks, but I wonder why the result of DN-Deformable-DETR in paper[1] is 43.4 and why 'Deformable' can bring so large improvement. DN-Detr-R50 get only 38.5 in paper[2] and I also verified it.

Mr-Bigworth commented 1 year ago

[1] DN-DETR: Accelerate DETR Training by Introducing Query DeNoising [2] GROUP DETR: FAST DETR TRAINING WITH GROUPWISE ONE-TO-MANY ASSIGNMENT

FengLi-ust commented 1 year ago

The 43.4 result is our initial implementation without deformable attention in the decoder and some other optimizations. Our released model performs better by better combine deformable detr and dn-detr.

DN-Detr-R50 gets only 38.5 because detection models without multi-scale features converge slowly and cannot do well on small objects. Multi-scale features are strong, and deformable attention is also strong to fuse multi-scale features well.

All these make it even stronger.

Mr-Bigworth commented 1 year ago

The 43.4 result is our initial implementation without deformable attention in the decoder and some other optimizations. Our released model performs better by better combine deformable detr and dn-detr.

DN-Detr-R50 gets only 38.5 because detection models without multi-scale features converges slowly and cannot do well on small objects. Multi-scale features are strong, and deformable attention is also strong to fuse multi-scale features well.

All this make it even more stronger.

Thanks very much for your answer!

hotcore commented 1 year ago

man i just wanna know how long you train your model on "dataloader.train.total_batch_size = 4 and train.max_iter = 360000" ? appreciate！！

Mr-Bigworth commented 1 year ago

man i just wanna know how long you train your model on "dataloader.train.total_batch_size = 4 and train.max_iter = 360000" ? appreciate！！

Training DN-Deformable-DETR-R50 (12 epoch) cost me about 55 hours on one Tesla A100

hotcore commented 1 year ago

man i just wanna know how long you train your model on "dataloader.train.total_batch_size = 4 and train.max_iter = 360000" ? appreciate！！

Training DN-Deformable-DETR-R50 (12 epoch) cost me about 55 hours on one Tesla A100

i use one v100 32g,same settings as you,detrex shows more than 4 days🤣 Anyway, thks for your reply!!