random seeds are now perturbed with time. and can now give random str…

[ffinger]

…eam to postfn

[jakeret]

By using time the entire system becomes non-deterministic i.e. it will be impossible to reproduce the exact same result.

[ffinger]

yes, I do agree. On the other hand I think it is quite important that the seed is not the same (1...N) in every calibration run. How about seeding the main process, drawing k random numbers (k=number of workers in the pool), which then serve to seed the generators of all workers? in this way the system is deterministic but still depends on the main seed, which can be changed.

also, I think it is very important to pass on the Random Stream to the postfn function, so that random numbers needed to compute the likelihood come from the same stream. otherwise they come from numpy.random, whose state random state (in Unix) is inherited from the main process, and thus the random numbers in the postfn function on different processes are identical.

[jakeret]

Ah, now I see the issue that the current reseeding causes. Creating the seeds in the main process is actually a good idea. I need to think about what happens when paralellizing with MPI but at the moment I think it could do the trick.

I would like to keep the function signature of postfn as simple as possible. One has the option to perturb numpy's seed using the passed theta values or time.

[ffinger]

I would like to keep the function signature of postfn as simple as possible. One has the option to perturb numpy's seed using the passed theta values or time.

How about making random state an optional argument then?

[jakeret]

Do you see an other way to do this than wrapping the postfn-call in a try-except? This is not very elegant, in particular because in 99% of the applications it will fall in the except block

[ffinger]

I do not have experience with this, but the inspect module might be an option. You can get the number of arguments a function takes as well as the names of kwargs.
http://stackoverflow.com/questions/4674875/python-function-parameter-count

[jakeret]

Inspecting the callable before the call to identify the number of params feels a bit like overkill and not very pythonic

[ffinger]

right. maybe use a parameter that can be True or False when you setup the
sampler object? it can be False by default, so no random state object is
given to postfn, but if it is True you give one.

2016-04-21 10:23 GMT+02:00 jakeret [email protected]:

Inspecting the callable before the call to identify the number of params
feels a bit like overkill and not very pythonic

—
You are receiving this because you authored the thread.
Reply to this email directly or view it on GitHub
#10 (comment)

[ffinger]

I just implemented what I wrote in my last comment. Up to you if you like it and want to include it in your version. For my application it is necessary.

[jakeret]

Could you give me some information on your application so that I can understand why it's neccessary?

[coveralls]

Coverage decreased (-0.6%) to 91.571% when pulling 3bad250 on ffinger:master into bd4e22f on jakeret:master.

[ffinger]

Because I am using it to calibrate a stochastic model (that means I make extensive use of random numbers in the model) and I would like to be able to control the rng state of the model too, otherwise the calibration is not reproducible.

[jakeret]

Can't you use the theta values passed to your function to seed the RandomState in your model?

[ffinger]

You mean a seed depending on the parameter values? I don't get it.

2016-04-22 10:45 GMT+02:00 jakeret [email protected]:

Can't you use the theta values passed to your function to seed the
RandomState in your model?

—
You are receiving this because you authored the thread.
Reply to this email directly or view it on GitHub
#10 (comment)

[jakeret]

Exactly, e.g. define the seed to be the sum or product over all the parameters. In principle the proposed parameters all come from a reproducible rng stream. It's not perfect but simple

[ffinger]

But then the seed is the same for a given parameter set. Not a good idea
for a stochastic model, where you want to have variability among different
runs of the same model with the same parameters.

2016-04-22 13:28 GMT+02:00 jakeret [email protected]:

Exactly, e.g. define the seed to be the sum or product over all the
parameters. In principle the proposed parameters all come from a
reproducible rng stream. It's not perfect but simple

—
You are receiving this because you authored the thread.
Reply to this email directly or view it on GitHub
#10 (comment)