results.txt interpretation

[jb-ssncse]

Hi,
1.Can you please tell, what are the values refer to after image size in results.txt while training
2. I can understand the working of detect.py, but Where to specify path for trained model on custom data in test.py ?

[glenn-jocher]

Columns in results.txt are losses (train and val) and metrics.

You pass --weights the same way in all python files, i.e.:
python test.py --weights /path/to/best.pt

[jb-ssncse]

Thank you for the prompt reply

[glenn-jocher]

BTW results.txt is plotted as results.png after training completes, I would recommend you look at results.png rather than trying to understand the txt file.

[jb-ssncse]

Thank you, i will look for the png file

[kashishgoyal31]

can you please suggest how can I analyse the results.png file for custom dataset. How can I get the explanation of the parameters. How do I know whether the training is file for do I need to retrain the model. Pls reply

[kashishgoyal31]

I want explanation of the graphs that are generated while training

[glenn-jocher]

@kashishgoyal31 results.png displays train and val losses and metrics (P, R, mAP)

[kashishgoyal31]

Could you please discuss in detail what does each of these graphs represent (box_loss, obj_loss, cls_loss)?
Difference between mAP_0.5 and mAP_0.5:0.95??

[glenn-jocher]

@kashishgoyal31 YOLO uses 3 loss components, box, objectness and classification. You can read about them in the YOLOv3 paper. https://arxiv.org/abs/1804.02767

For metrics simply do a Google search, i.e. 'object detection metrics'.

[kashishgoyal31]

I could not understand why the objectness loss increases in all cases in validation. What does obj_loss signify?

[kashishgoyal31]

Sir, I am facing another issue that when I am training on the custom dataset, the model is not accepting the images and prompts the error of a corrupted image, sharing a screenshot for the same. Please provide the solution as 90% of the dataset is not considered. Regards Kashish Goyal

…

On Mon, Aug 16, 2021 at 5:38 PM Glenn Jocher ***@***.***> wrote: @kashishgoyal31 <https://github.com/kashishgoyal31> YOLO uses 3 loss components, box, objectness and classification. You can read about them in the YOLOv3 paper. https://arxiv.org/abs/1804.02767 For metrics simply do a Google search, i.e. 'object detection metrics'. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#2191 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AVHEQRIJK3CZUDQMPCAZQ23T5D5VTANCNFSM4XQE34IQ> . Triage notifications on the go with GitHub Mobile for iOS <https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675> or Android <https://play.google.com/store/apps/details?id=com.github.android&utm_campaign=notification-email> .