Unable to replicate the results in DG-YOLO paper
karthikiitm87 opened this issue · 1 comments
karthikiitm87 commented
I am not sure why I am unable to replicate the results in DG-YOLO paper. I have used the given checkpoint and used the test.py code with the following command.
python test.py --weights_path weights/DGyolov3.pth --batch_size 1 --augment True
I got the following accuracies:
Average Precisions:
- Class '0' (echinus) - AP: 0.5152980911280326
- Class '1' (starfish) - AP: 0.20208827000334217
- Class '2' (holothurian) - AP: 0.23568195829279015
- Class '3' (scallop) - AP: 0.14365203861799805
- Class '4' (waterweeds) - AP: 0.001995890292044762
mAP: 0.21974324966684153
From the paper I see that mAP (in Table 3 for DG-YOLO) is 33.7. The only change I did is to impose the single batch-size. @mousecpn : Can you please help me if I am going terribly wrong some where. Why the direct pretrained model with a different batch_size should give significantly different results?
mousecpn commented
Maybe the problem is the environment. Try using pytorch1.0.0.