limit_train_batches

[sirtris]

Hi,
when I use the limit_train_batches flag in the trainer and set it to 0.1 does it use the same batches in each epoch or does it randomly sample 10% of the images (in my case) from the training data?

Thx

[akihironitta]

It completely depends on your dataloader's arg shuffle=True|False (or your sampler specifically).

[akihironitta]

So this means that when I set shuffle=True it loads different samples in every epoch, right? (I use the standard torch DataLoader, without any fancy sampler.)

@sirtris Yes, it should. Do you have any issue around it?

[sirtris]

Ok good to know. Maybe that could be clarified in the documentation.

I was using the argument to experiment with different train set sizes. But in the way it is currently implemented the experiments won't make much sense.

I think it would be great to have it implemented in a way such that only the same samples get sampled from the train set. Then one could also use the argument for experiments with reduced train set size.
Thanks

[sirtris]

So this means that when I set shuffle=True it loads different samples in every epoch, right?
(I use the standard torch DataLoader, without any fancy sampler.)

[rohitgr7]

yes, shuffle=True add a random-sampler for you.