Deprecate the "class" approach to the high-level kernels

[rlouf]

This was introduced to be able to access the low-level step and init functions at the higher level. With hindsight this was neither necessary nor a good design, and it also makes documenting these high-level constructs more complicated. We should use functions instead, like we do for window_adaptation.

[junpenglao]

Breaking: French dev propose to demolish class.

Joke aside, what is the proposal for high level API? Factory function that returns callables (init and step)?

[rlouf]

Yes. These classes currently serve no purpose other than proposing a shortcut to the lower-level step functions. Factory functions work just fine, and you can still import the lower level ones as e.g. blackjax.mcmc.hmc.step.

[albcab]

So the idea would be for the user to do

from blackjax import mcmc
state = mcmc.hmc.init(position)
step = mcmc.hmc.kernel()
new_state, info = step(rng_key, state, logdensity_fn, step_size, inverse_mass_matrix, num_integration_steps)

instead of

import blackjax
hmc = blackjax.hmc(logdensity_fn, step_size, inverse_mass_matrix, num_integration_steps)
state = hmc.init(position)
new_state, info = hmc.step(rng_key, state)

or do we want to introduce another function that does what classes do now? Maybe in blackjax.mcmc.hmc.py intead of kernels.py?

Also, this would work for VI, MCMC, SGMCMC, and SMC. But for adaptation algorithms we still need a higher level class or function that puts things together.

[rlouf]

I wasn't clear, you would still do:

import blackjax

hmc = blackjax.hmc(logdensity_fn, step_size, inverse_mass_matrix, num_integration_steps)
state = hmc.init(position)
new_state, info = hmc.step(rng_key, state)

except blackjax.hmc would be a function instead of a class.

[junpenglao]

So basically blackjax.hmc will just return a

blackjax/blackjax/base.py

Line 88 in 2f19d12

class MCMCSamplingAlgorithm(NamedTuple):

[albcab]

I see how having a class with a __new__ method that returns an MCMCSamplingAlgorithm object is redundant, when we can use a function to directly return that object, and if it makes documenting easier then refactoring is necessary. If all agree I can do the refactoring tomorrow, before new kernels are implemented as a class.

[rlouf]

Great, thanks! Then we can start moving these constructors to their respective algorithm-class folder like we discussed.

[albcab]

Notice that by replacing classes with functions we loose this alias: blackjax.mcmc.hmc.kernel, alias blackjax.hmc.kernel.

[rlouf]

I'm fine with that