tap --> 0.8 #2027
tap --> 0.8 #2027
Conversation
| @@ -27,7 +27,7 @@ function initialize_nuts(model::Turing.Model) | |||
| proposal = AdvancedHMC.NUTS{AdvancedHMC.MultinomialTS,AdvancedHMC.GeneralisedNoUTurn}(integrator) | |||
| adaptor = AdvancedHMC.StanHMCAdaptor( | |||
| AdvancedHMC.MassMatrixAdaptor(metric), | |||
| AdvancedHMC.StepSizeAdaptor(0.65, integrator) | |||
| AdvancedHMC.StepSizeAdaptor(0.8, integrator) | |||
There was a problem hiding this comment.
@JaimeRZP this shouldn't matter very much; if tests fail for 0.65, try to increase the number of samples.
Codecov ReportPatch and project coverage have no change.
Additional details and impacted files@@ Coverage Diff @@
## torfjelde/allow-abstractsampler-draft #2027 +/- ##
=====================================================================
Coverage 0.00% 0.00%
=====================================================================
Files 22 22
Lines 1462 1464 +2
=====================================================================
- Misses 1462 1464 +2
☔ View full report in Codecov by Sentry. |
| @@ -58,6 +59,11 @@ function AbstractMCMC.step( | |||
| # Link the varinfo. | |||
| f = setvarinfo(f, DynamicPPL.link!!(getvarinfo(f), model)) | |||
|
|
|||
| # Make init_params | |||
| if init_params === nothing | |||
| init_params = 4 .* rand(rng, LogDensityProblems.dimension(f)) .- 2 | |||
There was a problem hiding this comment.
Is this supposed to replicate the default initialization that already exists in DynamicPPL? The motivation is a bit unclear to me since 1) initial parameters will already be set using this approach by the code in DynamicPPL, 2) IIRC init_params is supposed to be given in the original untransformed space of the variables and hence this initialization might produce values outside of the support (in DynamicPPL this initialization is performed in the transformed space where these values are always valid!), and 3) init_params is not used below.
No description provided.