How to deal with the long-tailed problem in dataset,like ava data?

[ZJU-lishuang]

How to deal with the long-tailed problem in dataset?
Is it useful to simply increase the amount of data in the tail?

[innerlee]

For imbalanced data, you can

• change the sampling strategy, so that smaller classes are sampled more frequently, or
• change the sample weight in computing loss, or
• training specific models for the long-tailed subset

Which one is best is unknown

[ZJU-lishuang]

I will jsut try it.

[innerlee]

My recommendation is chaning the sampling. You can compute the median size of all classes, and in each epoch, randomly resample each class to have 2*median size samples (either subsampling or repeating, depending on the size of the class).

[irvingzhang0512]

WeightedRandomSampler is helpful. For now, WeightedRandomSampler is supported for single-gpu training.
Distributed WeightedRandomSampler is not supported in official pytorch yet.