Command line does not check that --init=0 provides a valid starting point about stan HOT 5 CLOSED

Yup, that's true. How should we handle it, warn the user and then exit
Stan? If that's cool I'll add a fix and make one of these new-fangeld
pull-requests.

On Tue, May 14, 2013 at 6:15 PM, Marcus Brubaker
[email protected]:

If the model fails with --init=0, Stan will run and appear to be working
fine, but will fail to do anything. We should check that all inits are
valid and error if a user specified init is broken.

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/41
.

from stan.

Yeah, I think that sounds right. We should make sure that the sampler is always starting from a valid initialization. Valid in this case meaning finite log prob and finite gradient. We check this already with random initialization, just need to do it for manually specified inits and bail with an appropriate error message if needed.

from stan.

Agreed -- just wanted to check that people didn't prefer switching to
random inits or something.

On Tue, May 14, 2013 at 6:46 PM, Marcus Brubaker
[email protected]:

Yeah, I think that sounds right. We should make sure that the sampler is
always starting from a valid initialization. Valid in this case meaning
finite log prob and finite gradient. We check this already with random
initialization, just need to do it for manually specified inits and bail
with an appropriate error message if needed.

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/41#issuecomment-17892707
.

from stan.

Ah, I see. I don't think we should switch to something else. If a user
specifies a particular initialization and that initialization isn't valid,
I think it's important that the user knows their initialization is broken.

Cheers,
Marcus

On Tue, May 14, 2013 at 1:49 PM, betanalpha [email protected]:

Agreed -- just wanted to check that people didn't prefer switching to
random inits or something.

On Tue, May 14, 2013 at 6:46 PM, Marcus Brubaker
[email protected]:

Yeah, I think that sounds right. We should make sure that the sampler is
always starting from a valid initialization. Valid in this case meaning
finite log prob and finite gradient. We check this already with random
initialization, just need to do it for manually specified inits and bail
with an appropriate error message if needed.

—
Reply to this email directly or view it on GitHub<
https://github.com/stan-dev/stan/issues/41#issuecomment-17892707>
.

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/41#issuecomment-17892908
.

from stan.

Pull request #74 fixed this issue.

from stan.