Merge ParticleSampler and HMCSampler into Sampler?

[yebai]

After introducing VarInfo and the new compiler interface, it seems possible to merge some sampler classes into a single one, e.g.

type Sampler{T:<InferenceAlgorithm}
  alg            ::  T
  particles   ::  ParticleContainer
  samples   ::  Array{Sample}     
  info        :: Dict{Symbol,Any} # Stores other sampler data, e.g. chunksize in HMC.          
end

In addition, to query sampler specific information (e.g. gradient sampler, particle sampler), we might consider introduce the following generic APIs for sampler

require_gradient(s :: Sampler) = false
require_particles(s :: Sampler) = false


# Then for HMC, we redefine these functions
require_gradient(s :: Sampler{HMC}) = true

# PG, SMC, IS
require_particles(s :: Sampler{PG}) = true
require_particles(s :: Sampler{SMC}) = true
require_particles(s :: Sampler{IS}) = true

[xukai92]

This makes sense. I'm also thinking the new sampler should somehow make it easier for different samplers to share some basic functions like assume(), rand(), etc. instead of using Union{}

[xukai92]

This data dict would also be helpful for adaptive samplers as we need to pass/store adaptive parameters somehow.

[yebai]

We should consider store run time information in sampler.info as well, e.g. acceptance ratio, time per MCMC iteration, time per leapfrog iteration etc.

[xukai92]

I think it's bit weird that all samplers should have a filed of particles :: ParticleContainer. I think we should treat it same as VarInfo. Do you think we can remove this field? I briefly looked into PG and SMC codes and feel it should be OK.

[xukai92]

@yebai I'm sorry that I pushed my commits (for this issue) directly to the master by mistake. Can you have a look at them?

[yebai]

It's OK. I will take a look and comment on the master branch.