Modifying the running dataset of the dataloader with a callback

[ysig]

Hi,

I'm modifying a pytorch lightning (1.4.2) code that uses trainer.fit(data, module).
I'm trying to apply a curriculum so that at batch_end, through a callback, every X amount of global steps I modify the dataset, which amounts in changing an attribute of the dataset:

• trainer.datamodule.train_dataloader().dataset.phase = phase #phase = 1
• trainer.datamodule.datasets['train'].phase = phase

but it turns out that in both ways the attribute modifications doesn't pass, which is probably because a shallow copy of the dataset is created. Do you know how I can access the actual dataset that is being running or how I can achieve the same goal in a different way?