Feature request: (re)start from a given position

[jae0]

Being able to (re)start as we can in the mcmc samplers? Or perhaps we can and I just need an example. Thx!

[Red-Portal]

Hi, thank you for your interest. I am not sure what you meant by MCMC sampler? AdvancedVI.jl doesn't involve MCMC.

[jae0]

Ah sorry. To be more precise, one can use "init_params = ..." to specify starting parameters in an mcmc run: sample( mymodel, Turing.NUTS(), 1000, init_params = starting.array) to start it from close to some previously (and quickly) estimated modes (say by MLE). I am dealing with slow convergence.

…

On Fri, Apr 11, 2025 at 9:03 PM Kyurae Kim ***@***.***> wrote: Hi, thank you for your interest. I am not sure what you meant by MCMC sampler? AdvancedVI.jl doesn't involve MCMC. — Reply to this email directly, view it on GitHub <#174 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AATUQOK5TWM6E37DP6DWUOT2ZBJ4XAVCNFSM6AAAAAB255MW72VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDOOJYGIZTMOBXGQ> . You are receiving this because you authored the thread.Message ID: ***@***.***> *Red-Portal* left a comment (TuringLang/AdvancedVI.jl#174) <#174 (comment)> Hi, thank you for your interest. I am not sure what you meant by MCMC sampler? AdvancedVI.jl doesn't involve MCMC. — Reply to this email directly, view it on GitHub <#174 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AATUQOK5TWM6E37DP6DWUOT2ZBJ4XAVCNFSM6AAAAAB255MW72VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDOOJYGIZTMOBXGQ> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[Red-Portal]

Ah I see! Thanks for the clarification. Which version of AdvancedVI with which optimizer are you using?

[jae0]

Oh! Onc can choose an optimizer too? How do I choose it? I have been using the default and only varying in ADVI(; adtype= ... ) My version seems to be: [b5ca4192] AdvancedVI v0.2.11 I have been trying Gaussian Process modelling via Cholesky and running into issues. Mostly related to non-positive definite matrices I think.

…

On Sat, Apr 12, 2025 at 9:39 AM Kyurae Kim ***@***.***> wrote: Ah I see! Thanks for the clarification. Which version of AdvancedVI with which optimizer are you using? — Reply to this email directly, view it on GitHub <#174 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AATUQONEIX6O2VTYZLUIWP32ZECPFAVCNFSM6AAAAAB255MW72VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDOOJYHAYTMNRUGU> . You are receiving this because you authored the thread.Message ID: ***@***.***> *Red-Portal* left a comment (TuringLang/AdvancedVI.jl#174) <#174 (comment)> Ah I see! Thanks for the clarification. Which version of AdvancedVI with which optimizer are you using? — Reply to this email directly, view it on GitHub <#174 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AATUQONEIX6O2VTYZLUIWP32ZECPFAVCNFSM6AAAAAB255MW72VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDOOJYHAYTMNRUGU> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[Red-Portal]

You can refer to the interface here. For the issue with Cholesky, there are lots of GP-specific tricks that you can use to prevent that. For example, adding a small "jitter" (e.g, 1e-7*I) to the covariance generally prevents minor numerical issues.

[jae0]

Ah yes, I see. I should have delved into the code. Thanks for the pointer and help!

…

On Sun, Apr 20, 2025 at 1:15 PM Kyurae Kim ***@***.***> wrote: You can refer to the interface here <https://github.com/TuringLang/Turing.jl/blob/fc32e10bc17ae3fda4d7e825b6fde45dc7bdb179/src/variational/advi.jl#L129>. For the issue with Cholesky, there are lots of GP-specific tricks that you can use to prevent that. For example, adding a small "jitter" to the covariance generally prevents minor numerical issues. — Reply to this email directly, view it on GitHub <#174 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AATUQOI7GQQOAX2BAQ3C2ZD22PBYLAVCNFSM6AAAAAB255MW72VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDQMJXGIZTQOBSG4> . You are receiving this because you authored the thread.Message ID: ***@***.***> *Red-Portal* left a comment (TuringLang/AdvancedVI.jl#174) <#174 (comment)> You can refer to the interface here <https://github.com/TuringLang/Turing.jl/blob/fc32e10bc17ae3fda4d7e825b6fde45dc7bdb179/src/variational/advi.jl#L129>. For the issue with Cholesky, there are lots of GP-specific tricks that you can use to prevent that. For example, adding a small "jitter" to the covariance generally prevents minor numerical issues. — Reply to this email directly, view it on GitHub <#174 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AATUQOI7GQQOAX2BAQ3C2ZD22PBYLAVCNFSM6AAAAAB255MW72VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDQMJXGIZTQOBSG4> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[Red-Portal]

This is now supported by specifying the state keyword argument in optimize.