search

lyuwenyu / RT-DETR

[CVPR 2024] Official RT-DETR (RTDETR paddle pytorch), Real-Time DEtection TRansformer, DETRs Beat YOLOs on Real-time Object Detection. 🔥 🔥 🔥
Apache License 2.0
2.61k stars 303 forks source link

RTDETR training batch size question #400

Open thisistejaspandey opened 3 months ago

thisistejaspandey commented 3 months ago

Hi, in the RTDETR paper, the network was trained for a batch size of 4 for 4 GPUs.

Why did you choose such a small batch size and would you expect better results with a larger batch?

Many thanks!

lyuwenyu commented 3 months ago
  1. because we have lots of p40 gpu; 😬
  2. If you want to accelerate the convergence of the model, you can try increasing bs