Use adaptive hmc #779
Draft
Use adaptive hmc #779
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Merge branch 'add-snaper-hmc' into adaptive-hmc-v2-i765 # Conflicts: # R/inference_class.R # tests/testthat/test_posteriors_geweke.R
…rror in trace_list_batches[[1]] : subscript out of bounds` - need to investigate further
…o do, but we are getting there!
Collaborator
Author
library(greta)
#>
#> Attaching package: 'greta'
#> The following objects are masked from 'package:stats':
#>
#> binomial, cov2cor, poisson
#> The following objects are masked from 'package:base':
#>
#> %*%, apply, backsolve, beta, chol2inv, colMeans, colSums, diag,
#> eigen, forwardsolve, gamma, identity, rowMeans, rowSums, sweep,
#> tapply
library(tictoc)
x <- normal(0, c(0.1, 1, 10, 100))
#> ℹ Initialising python and checking dependencies, this may take a moment.
#> ✔ Initialising python and checking dependencies ... done!
#>
m <- model(x)
tic()
draws_hmc <- mcmc(
model = m,
sampler = hmc()
)
#> running 4 chains simultaneously on up to 8 CPU cores
#>
#> warmup 0/1000 | eta: ?s warmup == 50/1000 | eta: 10s warmup ==== 100/1000 | eta: 5s warmup ====== 150/1000 | eta: 4s warmup ======== 200/1000 | eta: 3s warmup ========== 250/1000 | eta: 2s warmup =========== 300/1000 | eta: 2s warmup ============= 350/1000 | eta: 2s warmup =============== 400/1000 | eta: 2s warmup ================= 450/1000 | eta: 1s warmup =================== 500/1000 | eta: 1s warmup ===================== 550/1000 | eta: 1s warmup ======================= 600/1000 | eta: 1s warmup ========================= 650/1000 | eta: 1s warmup =========================== 700/1000 | eta: 1s warmup ============================ 750/1000 | eta: 1s warmup ============================== 800/1000 | eta: 0s warmup ================================ 850/1000 | eta: 0s warmup ================================== 900/1000 | eta: 0s warmup ==================================== 950/1000 | eta: 0s warmup ====================================== 1000/1000 | eta: 0s
#> sampling 0/1000 | eta: ?s sampling == 50/1000 | eta: 1s sampling ==== 100/1000 | eta: 1s sampling ====== 150/1000 | eta: 1s sampling ======== 200/1000 | eta: 1s sampling ========== 250/1000 | eta: 0s sampling =========== 300/1000 | eta: 0s sampling ============= 350/1000 | eta: 0s sampling =============== 400/1000 | eta: 0s sampling ================= 450/1000 | eta: 0s sampling =================== 500/1000 | eta: 0s sampling ===================== 550/1000 | eta: 0s sampling ======================= 600/1000 | eta: 0s sampling ========================= 650/1000 | eta: 0s sampling =========================== 700/1000 | eta: 0s sampling ============================ 750/1000 | eta: 0s sampling ============================== 800/1000 | eta: 0s sampling ================================ 850/1000 | eta: 0s sampling ================================== 900/1000 | eta: 0s sampling ==================================== 950/1000 | eta: 0s sampling ====================================== 1000/1000 | eta: 0s
toc()
#> 2.877 sec elapsed
par(mfrow = c(2, 4))
plot(draws_hmc, auto.layout = FALSE)tic()
draws_hmc_adapt <- mcmc(
model = m,
sampler = adaptive_hmc()
)
#> running 4 chains simultaneously on up to 8 CPU cores
#> sampling 0/1000 | eta: ?s sampling == 50/1000 | eta: 41s sampling ==== 100/1000 | eta: 20s sampling ====== 150/1000 | eta: 13s sampling ======== 200/1000 | eta: 9s sampling ========== 250/1000 | eta: 7s sampling =========== 300/1000 | eta: 5s sampling ============= 350/1000 | eta: 4s sampling =============== 400/1000 | eta: 4s sampling ================= 450/1000 | eta: 3s sampling =================== 500/1000 | eta: 2s sampling ===================== 550/1000 | eta: 2s sampling ======================= 600/1000 | eta: 2s sampling ========================= 650/1000 | eta: 1s sampling =========================== 700/1000 | eta: 1s sampling ============================ 750/1000 | eta: 1s sampling ============================== 800/1000 | eta: 1s sampling ================================ 850/1000 | eta: 0s sampling ================================== 900/1000 | eta: 0s sampling ==================================== 950/1000 | eta: 0s sampling ====================================== 1000/1000 | eta: 0s
toc()
#> 5.167 sec elapsed
plot(draws_hmc_adapt, auto.layout = FALSE)Created on 2025-03-12 with reprex v2.1.1
|
…t pass currently but this test driven development will be useful.
njtierney
force-pushed
the
adaptive-hmc-v3-i765
branch
from
8f96cf8 to
015e3d5
Compare
- remove unused n_adapt parameter in warm_up_sampler, instead refer to parameter set by user in adaptive_hmc() - removed unrequired sampler_param_vec in define_tf_kernel()
- Pass chains argument through to `check_geweke` as adaptive_hmc() requires at least 2 chains - use expect_no_error to test that adaptive_hmc() works with 0 warmup - update snapshots
…h as the free state, otherwise use the last warmed up parameters to pass through to sampling.
…is is implicitly checked in other tests, e.g., extra_samples etc
…c), and then for rejection (for counting bad samples)
- now returns as `self$warm_results`
…le time. - Remove extra `make_sampler_function()` - Add tests for `extra_samples` to ensure it works with `adaptive_hmc()`
… not the current state from the warmed up sampler
- replace `current_state` with `free_state` - warm_results doesn't need to return the kernel and current state - we actually just only need the kernel_results - make_sampler_function() doesn't need `sampler` arg - make_sampler_function() also doesn't need to have the logic inside of it to dispatch different results if it is already warmed up, we can move this inside of `sample_raw` - create sampling_results object that gets the result of `batch_results$kernel_results`, for use in geweke test
* Fix typos (proability and verbsoity) * add "as_tensor_spec" to globalVariables * update local snapshot tests (only works for me, but worthwhile checking
…it from "tune_tf". This doesn't do anything at the moment, but helps set things up for when we want to explore using tuning from the TF side of things and have slower equivalents in R that poll/check more frequently. Merge commit '6e1a347922086f993f6c75f7a672524c15555195' #Conflicts: # R/samplers.R
Collaborator
Author
|
Current progress - 10K warmup, 500 iterations:
Next steps:
Then for adaptive HMC geweke test
|
…or all four samplers
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Resolves #765