Modify dataloader at a given epoch

1

I need to switch on class-balanced sampling from a certain epoch. What is the most efficient way to achieve this?

I have tried creating a callback that sets the trainer.train_dataloader attribute inside the on_train_epoch_start hook. It checks the trainer.current_epoch and, if needed, tries to set the train_dataloader attribute to the appropriate Dataloader. However, during actual run, this step fails with an AttributeError saying that the attribute cannot be set.

The best alternative that I can think of right now is to redefine the train_dataloader in my DataModule. That is, choose a sampler based on self.trainer.current_epoch. And accompany this by setting the reload_dataloaders_every_n_epochs to some appropriate number.

Is there a better way to do this? I just need random sampling till, say epoch n, and class-balanced sampling thereafter.

2

I am gonna jump into this question because I want to know something kind of similar:
Is there a way to activate/deactivate a sampler between epochs? I would like to alternate if it is enabled or not.

3

Hi there

The best alternative that I can think of right now is to redefine the train_dataloader in my DataModule. That is, choose a sampler based on self.trainer.current_epoch . And accompany this by setting the reload_dataloaders_every_n_epochs to some appropriate number.

This is the way to do it, this is what the reload_dataloaders_every_n_epochs is meant for. And the same can be done in the LightningModule as well. Does it work for you?

4

Thank you for your confirmation. The technique does indeed solve the issue that I was trying to address.

1 Like