search

IDEA-Research / DN-DETR

[CVPR 2022 Oral] Official implementation of DN-DETR
Apache License 2.0
541 stars 62 forks source link

Batch size effects #16

Closed K-Mike closed 2 years ago

K-Mike commented 2 years ago

On my machine, I can only run a size 1 batch, how much will this degrade the results? I run with exactly the same parameters as yours the best one, except batch size, and the quality is much worse than MASK-RCNN Screenshot 2022-06-17 at 17 27 15 .

FengLi-ust commented 2 years ago

As we mentioned in ReadMe, our model is sensitive to batch size (like most DETR-like models). Our default setting is 16 batch size to reproduce the reported results.

image
K-Mike commented 2 years ago

Have you tried to accumulate gradient for multiple mini batches to simulate a larger batch size?

FengLi-ust commented 2 years ago

No, I have not tried this yet.