Hi @KumoLiu @ericspod ,

I remember the idea of _set_props is to save the instantiated python objects, so if some python object is not changed and you can initialize() again for some changed objects, it can reuse it. What do you think?

Thanks.

Hi @Nic-Ma is this being used somewhere or will be? If it is we should write a bit more commentary since the semantics wasn't so clear. If not and won't ever be used I would take it out. Thanks!

Hi @KumoLiu @ericspod ,
I remember the idea of _set_props is to save the instantiated python objects, so if some python object is not changed and you can initialize() again for some changed objects, it can reuse it. What do you think?
Thanks.

Hi @Nic-Ma is this being used somewhere or will be? If it is we should write a bit more commentary since the semantics wasn't so clear. If not and won't ever be used I would take it out. Thanks!

Hi @ericspod ,

There is a typical use-case: if we want to set the bundle idle on GPU and run inference for every input image, we can only override the data part and re-initialize the workflow, it can help avoid recreating the model and loading weights file to GPU.
@KumoLiu Can we add some comments in the example to explain?
What do you think?

Thanks.

Hi @ericspod,
Following an offline discussion with Nic, I've updated the tutorial. The self._set_props is utilized to store user-reset initialized objects, avoiding unnecessary re-initializations.

To aid understanding and also added as a note, I've provided a brief example.

wf = TrainWorkflow()
wf.initialize()
wf.run()
wf.loss = DiceFocalLoss(sigmoid=True)
# The workflow requires re-initialization, but the user-reset loss should remain as is.
wf.initialize()