I’m seeking help from cmdstan and I/O people on how to write num_warmup in csv output in a later stage, instead of during initialization. This came up when I experiment on terminating warmup dynamically so the actually num_warmup would be only available at the end of warmup. Intuition says a rewind…

Well, as long as its written before # Adaptation terminated # Step size = 0.822884 you are fine. And I guess this also finishes before adaptation (yes?). The following example returns the correct value if I do: a<- read_stan_csv(“~/Desktop/cmdstan/src/test/test-models/gq_model_output.csv”) a@si…

Well the definition of “num_warmup” changed. It’s now the maximum amount of warmup. I think this’d be the responsibility of whatever is parsing the output to figure out how much warmup is there. Presumably it’d just count the lines.

[image] bbbales2: I think this’d be the responsibility of whatever is parsing the output to figure out how much warmup is there. Presumably it’d just count the lines. we can either burden sampler stream writing or interface parsing, either works for me.

I would first remove the num_warmup line from the existing metadata written to the file and then add a similar line as we have for the mass matrix # "new warmup algorithm name" finished # num_warmup = 123 This should be easy to add. Let me know if you need help. So that would be # stan_version_…

[image] rok_cesnovar: I would first remove the num_warmup line from the existing metadata written to the file and then add a similar line Yeah that’s the easiest on Stan side, but requires changes on read_stan_csv, does it?

[image] bbbales2: Well the definition of “num_warmup” changed. What happens if different chains have different number of warmup iterations? It may wreak havoc downstream with interfaces. Nothing can get released until we know this won’t crash RStan, PyStan and CmdStan.

[image] Bob_Carpenter: What happens if different chains have different number of warmup iterations? I The way that the algorithm I’m implementing stops warmup guarantees same number of warmup iterations. Only that it’s no longer known a priori.

[image] yizhang: Only that it’s no longer known a priori . This may be a problem for RStan, which allocates ahead of time. It might be a problem with all the comments and the CSV reader. Our CSV format is very brittle.

[image] Bob_Carpenter: This may be a problem for RStan, which allocates ahead of time. It might be a problem with all the comments and the CSV reader. Our CSV format is very brittle. In that case there’s going to be even more rigging before cross-chain landing(if it can land).

[image] yizhang: if it can land I don’t think there’s any fundamental reason why it couldn’t. The changes shouldn’t be that hard. I think it’ll be worth it to get faster and/or more robust adaptation. Eventually we’ll have to deal with ragged chains. The only obstacle is that we have a lo…

Write dynamically determined number of warmup to csv output

Developers

yizhang February 8, 2020, 12:13am 9

The way that the algorithm I’m implementing stops warmup guarantees same number of warmup iterations. Only that it’s no longer known a priori.

Topic		Replies	Views
Minimizing warmup iterations - Error reading step size from CmdStan output Interfaces bug , cmdstanpy	4	575	August 15, 2023
New adaptive warmup proposal (looking for feedback)! Algorithms	50	4519	July 31, 2020
Saving & reusing adaptation in cmdstanr Interfaces cmdstanr	53	4335	June 8, 2022
Combine rstan::read_stan_csv's with different #s of samples? RStan	22	2512	June 23, 2017
Sampling fails after warmup CmdStan	20	1156	January 25, 2023

Write dynamically determined number of warmup to csv output

Related topics