Data generation when resuming from a checkpoint

[PC91]

Hi,
I have a question about the data generation when a training is resumed from a checkpoint.
In the training config file:

• The batch_type value is tokens.
• The seed value is fixed to make the result reproducible.
• The values of step and batches in the following line of trainer.py are observed (by printing to a logging file):

 for i, (batches, normalization) in enumerate(self._accum_batches(train_iter)):

            step = self.optim.training_step

I found that the value of step changes according to the checkpoint to resume. But this is not the case for batches. It seems that a training will always begin with the data generated from the first step even when this training is resumed from a checkpoint. Is there any way to change the configs to have a data generator that adapts to the checkpoint to resume ?

Thank you

[vince62s]

duplicate of #2006 ?
I recently added an index in the data processed. look here: #2496

actually I have to change the type (no need to be a tensor) but this can be used to achieve restoration of the proper index in datasets.
Do you want to work on this ?

[PC91]

duplicate of #2006 ? I recently added an index in the data processed. look here: #2496

actually I have to change the type (no need to be a tensor) but this can be used to achieve restoration of the proper index in datasets. Do you want to work on this ?

Thanks. I will have a look at it