Divergent wh (width-height) loss

[peterhsu2018]

Hello, i have a question about the hyperparameters setting in train.py,
the learning rate here and in my .cfg if different, what's the value will be used?

[glenn-jocher]

@peterhsu2018 your wh loss is diverging. You should increase your burnin period or decrease hyp['wh'], or alternatively implement a more stable wh loss as in #168.

[glenn-jocher]

To answer your other question the lr0 value in train.py is used, and the learning rate scheduler in defined in train.py as well:
scheduler = lr_scheduler.MultiStepLR(optimizer, milestones=[round(opt.epochs * x) for x in (0.8, 0.9)], gamma=0.1)

[peterhsu2018]

@glenn-jocher
I notice the optimizer
optimizer = optim.SGD(model.parameters(), lr=hyp['lr0'], momentum=hyp['momentum'], weight_decay=hyp['weight_decay']) will use it, and
scheduler = optim.lr_scheduler.MultiStepLR(optimizer, milestones=[218, 245], gamma=0.1, last_epoch=start_epoch-1) was command out, but the learning_rate from .cfg seems to not used?
The first issue i found the root cause is the filename of datasets, 'abc.png.7ed8sj.png' is the origin name,
after i replace to 'abc_7ed8sj.png' is fine.

[glenn-jocher]

@peterhsu2018 The learning rate from the .cfg file is only used if opt.plateau is False. To ensure correct usage, verify that opt.plateau in train.py is set to False. Regarding the dataset filename issue, thank you for sharing your solution. If you have any further questions or issues, feel free to ask!