Oversampling for imbalanced data

[YunchaoLiu]

Hi, I'm new to pytorch lightning. I'm trying to train on a very imbalanced dataset. I'm wondering if there is a way to do the oversampling with PL?

[rohitgr7]

PL doesn't take care of this. You can simply configure a weighted sampler and inject it to your datalaoder.
Ref: https://discuss.pytorch.org/t/how-to-handle-imbalanced-classes/11264/2

[daMichaelB]

Hey @rohitgr7 ,

i tried exactly that with creating a ImageClassificationData with the from_data_frame factory:

ImageClassificationData.from_data_frame(
            ...
            sampler=WeightedRandomSampler
        )

It seems that the behaviour in data_module._train_dataloader() function has changed and the sampler is initialized there:

I also found this in the docs:

It is expected to no longer pass a sampler instance, but a sampler class. I am wondering how i can pass all the parameters of the WeightedRandomSampler to this piece of code. I was not able to do it.

Should i open up a ticket or do you have any idea ? Thank you

[ethanwharris]

Hey @daMichaelB Mind opening an issue on the Flash repo for this? We should support both sampler type (in the event that it needs a dataset argument) / partials and instantiated samplers (for samplers like weighted random which don't need it).

Thanks 😃

[daMichaelB]

Hello @ethanwharris. Sure thing! i will open up an issue. Thanks for the quick response