Lasso and Glmnet library

Dear all,

I have a question about Lasso with a high-dimensional problem, both with the standard Lasso procedure in matlab and with the wrapper of the Glmnet Library. As to the standard procedure in Matlab (which solves the problem by means of ADMM algorithm), I can always get a solution which is sometimes less accurate than that produced by the wrapper of Glmnet library in Matlab. But, when I use the wrapper, I sometimes get this exception:

 Error using  == 
 Matrix dimensions must agree.

Error in cvglmnet (line 358)
idmin = options.lambda==CVerr.lambda_min;

Even if I explicitly declare the sequence of lambdas to be used, I still might get this exception. I have also seen that, in this case, the structure returned from the following lines code of function elnet (lines 23-24):

task = 11;
[lmu,a0,ca,ia,nin,rsq,alm,nlp,jerr] = glmnetMex(task,parm,x,y-offset,jd,vp,ne,nx,nlam,flmin,ulam,thresh,isd,weights,ka,cl,intr,maxit);

alm does not contain the sequence of lambdas required, but a vector of NaNs. What this might be due to? And what can I do to solve this problem?

Many thank for your attention and my best regards,

Jason.

[Le Zheng]

Hi Jason, I'm facing the same problem, too. Did you find a solution?

Answer 1

[Albena Vassileva]

EDIT: Figured it out. Or rather, "patched it".. The input should be in double precision. Still not sure exactly why it happened...

But I had it in single and was getting this error too. I can only assume it went wrong somewhere around the initialization of the lambda search, and that's why it was sometimes working.

[Le Zheng]

Hi Albena, I'm facing the same problem, too. But the error still comes out when I use cmd double to convert my data to double precision. Do you have any clues?