Add new sampler: weighted sampler

[marcoyang1998]

Add a weighted sampler, where each cut's sampling probability is proportional to its weight. This is useful for unbalanced dataset, where some classes have very few data. The weight for each cut should be computed by the user and passed to the sampler. It's similar to pytorch's WeightedRandomSampler (see here)

This sampler only works with eager manifest since we need to perform sampling globally.

[pzelasko]

Thanks @marcoyang1998, I appreciate your work. I think you could achieve a similar outcome by splitting the cutset into subset cutsets for each class, and then using mux to get a cutset to be passed to any of the existing samplers. It would also work with lazy manifests and bucketing.

class_cutsets = [cuts_class0, cuts_class1, ...]
class_weights = [w_class0, w_class1, ...]
cuts = CutSet.mux(*class_cutsets, weights=class_weights)

[marcoyang1998]

Hi Piotr,

thanks for the CutSet.mux example. I thought about this before as well, and it should work very well if the samples under the same class share an equal weight. However, in my setup, which is a multi-class classification (AudioSet), every sample has a unique sampling weight and using CutSet.mux for this scenario is impractical.

[pzelasko]

That's a valid point, I didn't think of this use-case. In that case LGTM, I just left one comment that may help us reduce code duplication.

[pzelasko]

I wonder whether this class can inherit SimpleCutSampler and only override the necessary parts (e.g. using WeightedDataSource in __init__); otherwise most of this code looks very similar.

[marcoyang1998]

Good idea! I just changed the inheritance, please have a look

[pzelasko]

Can we remove this assertion? let's just ignore the value of shuffle in this case.

[pzelasko]

is if self.shuffle branch needed here?

[pzelasko]

I made the changes I requested myself as I'd like to release a new version and include this. Thanks @marcoyang1998!