search

panlanfeng / KernelEstimator.jl

The julia package for nonparametric density estimate and regression
Other
26 stars 6 forks source link

installation issue #2

Closed PaulSoderlind closed 8 years ago

PaulSoderlind commented 8 years ago

Hi,

on Julia 0.4.5 (Windows 7, 64bit), Pkg.add("KernelEstimator") generates lots of error messages when trying to install Cubature.

However, this works: first install Cubature, then Yeppp and finally KernelEstimator.

Btw. the Require file says "julia 0.3." Wrong?

/Paul S

panlanfeng commented 8 years ago

Could you run 

Pkg.build("Cubature")

and see if there is any error?

This package itself can work both on julia 0.3 and 0.4. "Julia 0.3" means 0.3 and above.

PaulSoderlind commented 8 years ago

Hi again, I tried on two other machines. In these cases, Pkg.add("KernelEstimator") seems to work. However, at "using KernelEstimator" the pre-compile fails because of problems with Yeppp. Running Pkg.build("Yeppp") starts a download from/to a Yeppp-1.0.0\bindings folder. After that all works well.

Btw. could you add some more documentation on eg. the cross validation? Are you using an EPE as the loss function? My experience with EPE has led me to loop over different bandwitdths and then inspect the loss function graphically (non-monotic).

panlanfeng commented 8 years ago

Your previous problem with Cubature.jl is most likely a bug of Cubature or BinDeps. I will make a fix of the precompile issue soon. Currently you can get around that by restarting Julia and then you will be fine.

All the cross validation uses leave one out mean squared error, which can be found in most textbook or lecture notes about nonparametric, such as here http://www.ssc.wisc.edu/~bhansen/718/NonParametrics1.pdf In some cases cross validation does give bad bandwidth however there is no better alternatives. It is usually acceptable to adjust the bandwidth by a little bit if it gives a too small bandwidth.

If EPE means "empirical prediction error", then yes.

PaulSoderlind commented 8 years ago

Hi,

thanks.

uses leave one out mean squared erro that's what I meant by EPE, see eq 6.12 of the attached notes. See also Fig 6.6 for a result from looping over different bandwitdhs (in this case, it's well-behaved).

Thanks for providing the code.

/Paul S

On 05/04/2016 17:26, Lanfeng Pan wrote:

Your previous problem with Cubature.jl is most likely a bug of Cubature or BinDeps. I will make a fix of the precompile issue soon. Currently you can get around that by restarting Julia and then you will be fine.

All the cross validation uses leave one out mean squared error, which can be found in most textbook or lecture notes about nonparametric, such as here http://www.ssc.wisc.edu/~bhansen/718/NonParametrics1.pdf http://www.ssc.wisc.edu/%7Ebhansen/718/NonParametrics1.pdf In some cases cross validation does give bad bandwidth however there is no better alternatives. It is commonly acceptable to adjust the bandwidth by a little bit if it gives a too small bandwidth.

I am not sure what EPE means.

— You are receiving this because you authored the thread. Reply to this email directly or view it on GitHub https://github.com/panlanfeng/KernelEstimator.jl/issues/2#issuecomment-205857465

panlanfeng commented 8 years ago

Oops, I cannot see any attachment. If you replied this via email, an attachment will be removed by Github. You can send me email at pan@iastate.edu if you like.

panlanfeng commented 8 years ago

This should have been fixed since Cubature being updated.