mh_step() should take and return parameters as huge vector

[davidwhogg]

Basically the input and output of the mh code should be identical to the emcee code.

For example, the returned chain should be nsamples x nparameters in size / shape.

[davidwhogg]

I want the call

X_new = mh_step(X, inds, sigmas)

X is the current parameter vector inds is a list of integers saying which of the parameters should be sampled in this step sigmas is a list of root-variances of the gaussians for the proposal distribution for the parameters listed in inds

This function will try to take a step just in the parameters listed in inds.

If you iterate over choices for inds, you will be Gibbs sampling. Does that make sense?

[julywater]

I thought about this way. But the inconvenient part is that while sampling Gamma in patch i only data in patch i will be used. while sampling P(shape parameter ) all the data will be used. I don't know how to use a general function to do that yet.

2012/12/20 David W. Hogg notifications@github.com

I want the call

X_new = mh_step(X, inds, sigmas)

X is the current parameter vector inds is a list of integers saying which of the parameters should be sampled in this step sigmas is a list of root-variances of the gaussians for the proposal distribution for the parameters listed in inds

This function will try to take a step just in the parameters listed in inds.

If you iterate over choices for inds, you will be Gibbs sampling. Does that make sense?

— Reply to this email directly or view it on GitHubhttps://github.com/julywater/WeakLensing/issues/5#issuecomment-11560563.

[julywater]

Oh,maybe I can do it by some change in the postfunction(),I'll think about it tomorrow.

2012/12/20 yike tang tangyike455665@gmail.com

I thought about this way. But the inconvenient part is that while sampling Gamma in patch i only data in patch i will be used. while sampling P(shape parameter ) all the data will be used. I don't know how to use a general function to do that yet.

2012/12/20 David W. Hogg notifications@github.com

I want the call

X_new = mh_step(X, inds, sigmas)

X is the current parameter vector inds is a list of integers saying which of the parameters should be sampled in this step sigmas is a list of root-variances of the gaussians for the proposal distribution for the parameters listed in inds

This function will try to take a step just in the parameters listed in inds.

If you iterate over choices for inds, you will be Gibbs sampling. Does that make sense?

— Reply to this email directly or view it on GitHubhttps://github.com/julywater/WeakLensing/issues/5#issuecomment-11560563.

[davidwhogg]

I don't think it matters that you are using only part of the data. The passing of data is not what is making the code slow.

[julywater]

I wrote a gibbs sampler based on a general mh_step() function as you said. Then I find a big problem I can't get rid of the loop for i in xrange(NP) (MH_step() for each patch) All the possible way I came up with have conflicts between sampling gammas and alpha...

Then I thought It may be better to write different mh_sampler for gamma and alpha sampling as I did last time. It's more direct and won't be slower.

in https://github.com/julywater/WeakLensing/blob/master/code/MH.py I have both versions

[davidwhogg]

A for loop that goes over a big piece of code is not as bad as an "inner loop". But if you can do the MH loop with a map() operation, that's good because then we can use multiprocessing to parallelize it.