About best practices for implementing loss functions

[minasora]

Thank you for your excellent work. I want to ask for the best practices for the loss function. I have seen in the Get Start, the loss function is in the model class:

        if mode == 'loss':
            return {'loss': F.cross_entropy(x, labels)}

How should I separate the loss function and set it in the config just like mmdet?

   loss_cls=dict(
            type='CrossEntropyLoss', use_sigmoid=True, loss_weight=1.0),

What is the best practice for the position of the loss function in mmengine? I searched the documentation but couldn't find an answer. Could you please enlighten me?

[narthchin]

Take an example from mmdet:
https://github.com/open-mmlab/mmdetection/blob/main/mmdet/models/dense_heads/anchor_head.py
Similarly, you can add loss_cls and more in your model __init__ func and use MODELS.build to build the loss func from your config file, then in your forward function or somewhere else:

losses_cls = self.loss_cls(pred, gt)
return dict(loss_cls=losses_cls)

[minasora]

Thank for your reply!