DataLoader but with strings not in GPU memory?

[turian]

I have a DataLoader reading from a DataSet, where each instance is (x, y, string).

x and y are tensors I want loaded to GPU memory, which is great because pl does that for me.

However, I don't want the string in GPU memory, since it is only used for crossreferencing another table for scoring.

Or is there a workaround that I can use to do my string lookups?

[tchaton]

Hey @turian,

Only torch.tensor are moved to the GPU. A string won't be moved, so I am not sure to fully follow your questions.

Would you mind to share a script to better explain your problem ?

Best,
T.C

[turian]

@tchaton basically, I want to maintain an index or a lookup to auxiliary information for each instance. What is the cleanest pattern for doing so?

Thank you

[tchaton]

The cleanest to me, would be to override the default_collate to make sure the indices aren't transformed into tensors and therefore aren't move automatically by Lightning on the device.

We already have similar mechanism in Lightning Flash: https://github.com/PyTorchLightning/lightning-flash

[turian]

@tchaton do you mind pointing me to which file that's in?

Do you mean:

https://github.com/PyTorchLightning/lightning-flash/blob/a1bece4361a7cb5449715bef9975696e96c5f9ae/flash/core/integrations/vissl/transforms/utilities.py

[tchaton]

Dear @turian,

Are you using Lightning Flash or another framework ?

If you are using your own code and DataLoader creation, then you can do DataLoader(ds, collate_fn=my_custom_collate_fn), otherwise you can create an issue on Flash with a reproducible script there and we would help you out !

Best,
T.C

[turian]

@tchaton I am just using lightning, not flash.