Support of ReverseDiff.jl as AD backend #428
Conversation
| @@ -0,0 +1,14 @@ | |||
| using Turing: VarEstimator, add_sample!, get_var | |||
There was a problem hiding this comment.
@yebai Unit test for var estimator is here.
src/samplers/support/adapt.jl
Outdated
| update_state(wum::WarmUpManager) = begin | ||
| wum.iter_n += 1 # update iteration number | ||
| # Ref: https://github.com/stan-dev/stan/blob/develop/src/stan/mcmc/hmc/nuts/adapt_diag_e_nuts.hpp | ||
| update_pre_cond(wum::WarmUpManager, θ_new) = begin |
There was a problem hiding this comment.
change update_pre_cond to update_pre_cond! since it modifies wum.
There was a problem hiding this comment.
do we only support diagonal preconditioning matrix at the moment?
| @@ -10,7 +11,7 @@ end | |||
|
|
|||
| sample_momentum(vi::VarInfo, spl::Sampler) = begin | |||
| dprintln(2, "sampling momentum...") | |||
| randn(length(getranges(vi, spl))) .* spl.info[:wum][:stds] | |||
| randn(length(getranges(vi, spl))) ./ spl.info[:wum][:stds] | |||
There was a problem hiding this comment.
@yebai how to use pre-cond (1)
src/samplers/support/hmc_core.jl
Outdated
| @@ -30,13 +35,35 @@ leapfrog(_θ::Union{Vector,SubArray}, p::Vector{Float64}, τ::Int, ϵ::Float64, | |||
| p_old = p; θ_old = copy(θ); old_logp = getlogp(vi) | |||
|
|
|||
| p -= ϵ .* grad / 2 | |||
| θ += ϵ .* p # full step for state | |||
| θ += ϵ .* p .* (spl.info[:wum][:stds].^2) # full step for state | |||
There was a problem hiding this comment.
@yebai how to use pre-cond (2)
There was a problem hiding this comment.
there is an error here, the correct way should be
θ += ϵ .* p
the current code effectively cancelled pred-cond(1) operation.
There was a problem hiding this comment.
why does Stan do this? I made these changes mainly by referring https://github.com/yebai/Turing.jl/blob/1df604591e4e4297261dfcbb478ca51059ecea9d/src/samplers/support/hmc_core.jl
src/samplers/support/adapt.jl
Outdated
| type WarmUpManager | ||
| iter_n :: Int | ||
| state :: Int | ||
| adapt_n :: Int | ||
| params :: Dict |
There was a problem hiding this comment.
Use a more concrete type for params :: Dict if possible to improve efficiency.
| @@ -52,35 +79,58 @@ find_H(p::Vector, model::Function, vi::VarInfo, spl::Sampler) = begin | |||
| # This can be a result of link/invlink (where expand! is used) | |||
| if getlogp(vi) == 0 vi = runmodel(model, vi, spl) end | |||
|
|
|||
| p_orig = p ./ spl.info[:wum][:stds] | |||
| p_orig = p .* spl.info[:wum][:stds] | |||
There was a problem hiding this comment.
@yebai how to use pre-cond (3)
There was a problem hiding this comment.
this is correct; perhaps we can change p_orig to p_prime
p_prime = p .* spl.info[:wum][:stds]
H = dot(p_prime, p_prime) / 2 + realpart(-getlogp(vi))
src/samplers/support/adapt.jl
Outdated
| wum[:stds] = ones(D) | ||
| wum[:vars] = ones(D) | ||
|
|
||
| wum[:stds] = 1.0 |
|
@yebai I tagged you a few places in the code. For adaptation in general, codes are mostly in |
src/samplers/support/adapt.jl
Outdated
|
|
||
| spl.info[:wum] = wum | ||
| end | ||
|
|
||
| update_da_params(wum::WarmUpManager, ϵ::Float64) = begin | ||
| wum[:ϵ] = [ϵ] | ||
| reset_da(wum::WarmUpManager) = begin |
There was a problem hiding this comment.
perhaps change reset_da to reset_da!
src/samplers/support/adapt.jl
Outdated
| end | ||
|
|
||
| return false | ||
|
|
||
| end | ||
|
|
||
| adapt(wum::WarmUpManager, stats::Float64, θ_new) = begin |
src/samplers/support/adapt.jl
Outdated
| # Dual averaging | ||
| wum[:ϵ] = [] | ||
| reset_da(wum) | ||
| wum[:n_warmup] = spl.alg.n_adapt |
There was a problem hiding this comment.
perhaps merge spl.alg.n_adapt and spl.alg.delta into spl.info[:adapt_conf]?
|
@yebai Anything left to do in this PR? |
* Fix dep log in lad * Dont send opt res * Fix VarInfo.show bug * Fix auto tune * Change * to .* in leapfrog * temp fix type * Disable @suppress_err temporarily * Fix a dep * Workable ReverseDiff v0.1 done * Add RevDiff to REQUIRE * Fix bug in R-AD * Fix some bugs * Fix bugs * Update test * ReversedDiff.jl mutable bug fixed * Any to Real * update benchmark * Resolve mem alloc for simplex dist * Fix bug and improve mem alloc * Improve implementaion of transformations * Don't include compile time in benchk * Resolve slowness caused by use of vi.logp * Update benchmark files * Add line to load pickle * Bugfix with reject * Using ReverseDiff.jl and unsafe model as default * Fix bug in test file * Rename vi.rs to vi.rvs * Add Naive Bayes model in Turing * Add NB to travis * DA works * Tune init * Better init * NB MNIST Stan added * Improve ad assignment * Improve ad assignment * Add Stan SV model * Improve transform typing * Finish HMM model * High dim gauss done * Benchmakr v2.0 done * Modulize var estimator and fix transform.jl * Run with ForwardDiff * Enable Stan for LDA bench * Fix a bug in adapt * Improve some code * Fix bug in NUTS MH step (#324) * Add interface for optionally enabling adaption. * Do not adapt step size when numerical error is caught. * Fix initial epsilon_bar. * Fix missing t_valid. * Drop incorrectly adapted step size when necessary (#324) * Edit warning message. * Small tweaks. * reset_da ==> restart_da * address suggested naming * Samler type for WarmUpManager.paras and notation tweaks. * Bugfix and adapt_step_size == > adapt_step_size!
TODOs