How to reproduce the results

[yurkovak]

Hi! Thank you for your work. I'm trying to reproduce your results and I get the following mismatch:

Model	Resolution	mAPval 0.5:0.95	mAPval 0.5:0.95 reproduction
YOLOv6-n	640	35.0	34.7
YOLOv6-tiny	640	41.3	41.4
YOLOv6-s	640	43.1	42.6

As you can see, Nano and Small models don't reproduce. I used

python -m torch.distributed.launch --nproc_per_node 4 tools/train.py --batch 256 --conf configs/yolov6n.py --data data/coco.yaml --device 0,1,2,3

for Nano (4, GPUs) and

python -m torch.distributed.launch --nproc_per_node 8 tools/train.py --batch 256 --conf configs/yolov6s.py --data data/coco.yaml --device 0,1,2,3,4,5,6,7

for Small (8 GPUs).

In both cases, this is the batch size and configs you mention in the README. Is there something else I should do to get your result? Thanks in advance!

[shensheng272]

Hi, thanks for reproduce exp. You might try --batch 128 for n and s. 0.2~0.5% gap is normal because of random seed.

[hhaAndroid]

I train with the same result. mAP is 42.6

[shensheng272]

New models and scripts for reproducing train/val result are available now. More info in readme. By the way, we publish result on different test scheme comparing to yolov5/yolov7. Still you could compare eval result in training whose test scheme is same to yolov5/yolov7. AP result of two test scheme actually quite close to each other .