limit_train_batches training

[roman-vygon]

Hi everyone!
I've been dealing with this memory leak and couldn't find a good solution, so I decided to reduce the epoch size of my dataset (since memory clears every epoch).
I've set limit_train_batches to 0.1 and increased the number of epochs 10 times. Is that OK?
Is the dataset shuffled before choosing the 10% for the next epoch or not?
I'm assuming that 10 epochs with limit_train_batches=0.1 is almost the same is 1 epoch with limit_train_batches=1.0, with only difference being that items can be randomly chosen multiple times.

[justusschock]

The dataset is shuffled when you create the dataloader/the sampler. So when you specify to reload the loaders for each epoch, you should be fine :)