Add SyntheticStream and ClassSampler for experiments reproduction

Written with flexibility in mind

This adds two main classes - SyntheticStream and ClassSampler. Example use-cases are implemented in tests.

ClassSampler

This is a building block of SyntheticStream. It works with python next() interface to provide easy way to iterate over samples. There are two ways the sampler will stop iterating:

In constructor we defined max_samples argument, and next sample will overload this limit
- In such case, ClassSampler raises StopIteration exception, which is compliant with Python iter iterface
- to check weather calling next() will raise StopIteration (this is used inside SyntheticStream), ClassSampler has end_of_iteration property, which returns bool value
There are no more samples to get, and ClassSampler was initiated with eoc_strategy equal to raise
- In such case, ClassSampler with raise EndOfClassError
- This behavior is not handled inside SyntheticStream - it should be user responsibility to make sure class samples will not end in experiment

SyntheticStream

It works as a aggregation of ClassSampler. Every class in the stream should have its own unique sampler (there can be no active duplicates - to avoid undefined behavior, but we can "hook" 2 samplers from 2 classes in different stream positions)

For every sampler, we specify weight_func(t) which returns importance of a class in stream at moment t.

Between every iteration we can check normalized probabilities of all classes in the stream (past, current and future - from t point of view) using class_probabilitites property. It is calculated inside next() transition, so output is result from last transition - i.e. accessing it after first iteration would give probabilities at time 0.

Future improvements

Currently ClassSampler loads samples as a python list which could lead to problems with huge datasets

rswc / ml-ids