Yolov3: Understanding the training output

[Arunavameister]

Can someone please help me understand the training output?
I understand some of the terms but not all.
Thank you.

[glenn-jocher]

GIoU is the box regression loss, wh is zero because GIoU is used instead, obj is objectness loss, cls is class loss, total is the total loss. targets is the number of targets (objects) per batch, and img_size is the pixels of the long side of the images.

The second set of metrics is the evaluation of the current network on the validation set. Your mAP is 0.21 on the 114 images you tested after 699 epochs.

[Arunavameister]

Thank you for the prompt answer!
Are there any threshold values(of sorts) that one can be on the lookout for? For example the intersection of union metric, what should be an optimal value where i can stop training?!
Same with the total loss
P.S - This repo is very useful for a lot of object detection tasks. Thanks for making it open source!

[glenn-jocher]

@Arunavameister you should plot your training with from utils import utils; utils.plot_results() to understand how the losses and metrics are evolving over time.