Manual iteration over dataloader memory leak, how to store the iterator correctly?

[Oliver]

Hi everyone,

I’ve been experiencing this issue:

Memory leak while iterating over dataloader

In an ideal case I prefer enumerating over dataloader with a for loop. But for a particular implementation, I need multiple batches within one iteration in a sequence. To do this I use, next(iter(dataloader)) to get as many batches as I need within one iteration. The issue is, this causes linear increase in RAM over iterations and epochs. Somehow, the memory is not cleared after every iteration. Any way I can avoid using iter and get multiple batches within one iteration?

From what I understood I should do is store the iterator which I do.
I.e. iterds = iter(dataloader)

then I get the item with next(iterds)

Is this also using the cycle() method that is described by the OP of the old issue?

How would I do it correctly then?

Many thanks in advance!

[Oliver]

Did anyone have an answer to this?

[ptrblck]

It’s unclear what your exact use case is, so could you post a minimal and executable code snippet reproducing the issue?