qiang chen

Ok, I will try it.

Multi-scale training is supported by Detectron2. You can refer to [this repo](https://github.com/hyz-xmaster/swa_object_detection) for swa. The results for the multi-scale training and saw are not included in this repo. You can...

You may need to re-train YOLOF with small image sizes. The provided pre-train model is trained with relatively large image sizes (from 512 to 768), which is not suitable to...

Hi, we do not have a plan to provide the 3-rd stage model now. Maybe next month. I'll keep this issue open. You can train the third stage on your...

You can check the caption of the Table. When comparing with YOLOv4, we first fuse conv and bn, then keep `cudnn_benchmark` as True, and use half-precision for inference.

Ok. The code will be refactored in the next few days.

Will do it tomorrow.

Please check this pr (#13) for progress.

YOLOF does not use SyncBN by default. This error is caused by [the calculation of `num_foreground`](https://github.com/megvii-model/YOLOF/blob/main/playground/detection/coco/yolof/yolof_base/yolof.py#L210). You can replace it with [the new version](https://github.com/chensnathan/YOLOF/blob/master/yolof/modeling/yolof.py#L396) to support training with one GPU.

You can specify GPU ids with `CUDA_VISIBLE_DEVICES`. For example `CUDA_VISIBLE_DEVICES=4,5,6,7 pods_train --num-gpus 4`, it will use the last 4 GPUs for training. **You may need to adjust the warmup iterations...