Skip metric regularization if adapt window=0 #3037
Conversation
Jenkins Console Log Machine informationProductName: Mac OS X ProductVersion: 10.11.6 BuildVersion: 15G22010CPU: G++: Clang: |
|
Although I largely agree with the update technically inverse metric updating is skipped for both adapt_window = 0 and adapt_window = 1. I'll change the pull name before merging. |
| covar = (n / (n + 5.0)) * covar | ||
| + 1e-3 * (5.0 / (n + 5.0)) | ||
| * Eigen::MatrixXd::Identity(covar.rows(), covar.cols()); | ||
| if (estimator_.num_samples() > 1) { |
There was a problem hiding this comment.
Because this behavior probably won't be obvious I think we need to wire a message to the logger. Both here and in var_adaptation can you add a std::string& msg argument which passes the message "Warning: inverse metric is not updated for window sizes less than 2". Then we can pass that to the logger in all of the transition implementations.
| @@ -27,3 +27,32 @@ TEST(McmcCovarAdaptation, learn_covariance) { | |||
| } | |||
| EXPECT_EQ(0, logger.call_count()); | |||
| } | |||
|
|
|||
| TEST(McmcCovarAdaptation, learn_covariance_one_sample) { | |||
There was a problem hiding this comment.
Corresponding test in var_adaptation_test.cpp?
|
If changes are made according to #3027 (comment) then no warning message would be needed for the base_window = 0 case, in which case messages wouldn't have to be passed, but if inverse metric updating is also disabled for base_window = 1 then I think we still need the message. |
Submission Checklist
./runTests.py src/test/unitmake cpplintSummary
This came up in #3027 (comment)
CmdStan accepts
window=0as a valid argument even though it means the adaptation cannot possibly estimate the inverse metric. The intended effect appears to be adapt only stepsize and leave the metric at its initial value, either the default identity or a user specified value. That works but only if you also setinit_buffer=0. Ifinit_buffer>0the sampler tries to update the metric anyway and messes it up.The metric update code is
stan/src/stan/mcmc/var_adaptation.hpp
Lines 24 to 28 in 35a15c8
The Welford estimator's
sample_variance()is a no-op if there are not enough samples so that part's fine.The problem is the subsequent regularization which in case of zero samples just discards the (unchanged) metric.
The fix is to simply skip the regularization when
sample_variance()didn't do anything.Intended Effect
Setting
window=0always keeps the initial metric.How to Verify
Run the
examples/bernoullimodel withadapt init_buffer=10 window=0and look for the mass matrix in output.csv.On develop it says
On this branch
Side Effects
Documentation
Copyright and Licensing
Please list the copyright holder for the work you are submitting (this will be you or your assignee, such as a university or company): Niko Huurre
By submitting this pull request, the copyright holder is agreeing to license the submitted work under the following licenses: