Using a fixed step size for HMC (or NUTS) and other sampler options

mcmc

#1

Is it possible to completely fix the stepsize__ parameter? Equivalently I tried to run the HMC sampler with warmup=0 but I got flat lines and zero-likelihood/negative poisson parameter warnings even when setting the initial parameters to the “true” values. Is that not allowed in Stan?

I was under the impression that with NUTS it may be necessary to have some optimization during the burn-in period, but for HMC only the step size required optimization, so it did not need anything else if a step size (as well as int_time was provided).

Additionally, since I’m asking about sampler options, is there one to suppress all warnings during sampling? verbose=False doesn’t seem to eliminate all messages, and for long runs I would like to see only the percent progress.

Thanks.


#2

can anyone help with this, just yes or no on whether stepsize__ can be fixed all the way through or sampler run with warmup=0 ? Thanks in advance.


#3

Sorry for the delayed response. This should be possible in cmdstan with the engine=static option. Is that not working for you?

Besides timestep, Stan adapts what is called its Euclidean metric during warmup. In an ideal case this should be something like the curvature of parameter space near where the posterior mass is. If this isn’t initialized, sampling can be super inefficient and fail unless the timestep is super small.


#4

No problem. I’m surprised you guys are able to reply to nearly all questions, I was just afraid it would be lost at the bottom of the scrolling.

I am using PyStan, is that not feasible with that interface (or Julia, or even R, I’ve never used cmdstan, but would be comfortable with a few others if necessary). I am having trouble with parallel chains ending up with different parameters at optimization for large inference problem, that’s why I wanted to fix static HMC’s parameters from previous long runs. I forgot about the mass matrix, but I didn’t see its parameters on the sampler parameters, maybe I missed that.

Does this all make sense?
Thanks for the attention.


#5

Aaah, if you want to set the mass matrix too then you will need to use cmdstan. I think it’s the only interface at this point that supports that unfortunately. Also you might need to use cmdstan to get the mass matrix from an already-adapted chain. I don’t think it’s available in Rstan and I’m not sure about Pystan.


#6

Thanks, that’s very helpful. It would be some work to get all the python input adapted to cmdstand, but I may try to go for it if it turns out reducing the number of parameters to be tuned by switching to HMC doesn’t help. Thanks again.


#7

cmdstan accepts input in JSON format now.


#8

to use JSON input for cmdstan, you’ll need to current (develop branch) of cmdstan and stan. if the data input file ends in .json, it will be parsed as such. ditto input files for specified initial mass matrix.

from the cmdstan manual:

Stan follows the JSON standard. A Stan input file in JSON notation consists of JSON object which contains zero or more name-value pairs. This structure corresponds to a Python data dictionary object. The following is an example of JSON data for the simple Bernoulli example model:

` { "N" : 10, "y" : [0,1,0,0,0,0,0,0,0,1] }`

Matrix data and multi-dimensional arrays are indexed in row-major order. For a Stan program which has data block

data { int d1;
       int d2;
       int d3;
       int ar[d1, d2, d3];
}

the following JSON input file would be valid

     { "d1" : 2,
       "d2" : 3,
       "d3" : 4,
       "ar" : [[[0,1,2,3], [4,5,6,7], [8,9,10,11]],
               [[12,13,14,15], [16,17,18,19], [20,21,22,23]]]
}

JSON ignores whitespace. In the above examples, the spaces and newlines are only used to improve readability and can be omitted.

All data inputs are encoded as name-value pairs. The following table provides more examples of JSON data. The left column contains a Stan data variable declara- tion and the right column contains valid JSON data inputs.


#9

May be worth the extra work to get additional control over the MCMC parameters. Thanks.