Is there a need to restore the dataloader condition

[ryanxingql]

Hi,
I am wondering that is there a need to restore the dataloader condition.
Every time we restore the model, we only restore the model parameters, optimizer, and scheduler.
The dataloader will start from the beginning, resulting in repeated training over the data in front.
To actually restore the training condition, we may need to restore the dataloader condition, i.e., skip the trained iters or epochs of data.
If this function is necessary, I may try to implement it.

[zengyh1900]

Hi @ryanxingql , that's an interesting question!
In my understanding, the data batch from the dataloader is typically shuffled. That means when we resume training, the data batch will be randomly selected from the whole dataset, rather than 'the front'.

[ryanxingql]

Thank you for your response. ^ ^

[ryanxingql]

If we set a seed for the data loader (i.e., for the data sampler), the data batch will not be randomly shuffled. You understanding may be correct when no seed is set.

[zengyh1900]

If we set a seed for the data loader (i.e., for the data sampler), the data batch will not be randomly shuffled. You understanding may be correct when no seed is set.

You're right!

[zengyh1900]

@ryanxingql We appreciate your passion for contributing MMEditing!
Thanks a lot 🤩 !

We found some of your projects are really interesting!
https://github.com/ryanxingql/powervqe
https://github.com/ryanxingql/winner-ntire22-vqe
https://github.com/ryanxingql/stdf-pytorch

And I have read your paper! https://github.com/ryanxingql/rbqe That's really an interesting idea!

Would you like to contribute these works to MMEditing?
Please don't hesitate! We're here to help if you need it!

[ryanxingql]

Thank you so much for your kind words ^ ^ I will contribute to this repo in my spare time.