jiayuzhou / malsar Goto Github PK

Are you planning to release an R package similar to this
MATLAB library?

Excellent work!
When i performed MALSAR/functions/sparse_low_rank/Least_SpareTrace.m, I found a missing function.

Undefined function or variable 'eplb'.

Error in Least_SparseTrace (line 171)
        [s_hat, lambda_new ]=eplb(s_bar, size(s_bar, 1), tau, lambda_old);

I look around the MALSAR file, and I can't find that function.
Can you provide that function? or tell me how to solve that problem.
Thank you so much!
Excellent work again! i'm so admire.

Hi jiayu:

I added the Logistic_Dirty.m and Logistic_rMTL.m in my repo, and they are tested using multiple gene expression data-sets.
Let me know if you want to merge it into your repo.

https://github.com/hank9cao/MALSAR/tree/master/MALSAR/functions/dirty
https://github.com/hank9cao/MALSAR/tree/master/MALSAR/functions/rMTFL

@jiayuzhou BTW, I noticed you guys want to write the python version of the MALSAR, how did you want to do it? To be honest, MALSAR will get more R users than python users, since R is more common laboratory language across many scientific fields(biology, geography) which are more interested in MALSAR.

Regards,
Hank

Hi Jiayu:

I was testing the proximal operator for the L1inf norm which lead me to test the prf_lbm.cpp.
First, it worked well, and the procedure to make sparsity is quite similar to the procedure in Sparse PCA.
But I found some source codes doesn't match the performance, here is my testing:
D=[1 2 2; 4 4 3];
tau=1;
[mu, theta , mc]=prf_lbm(D,2,3,tau);
Then

mu =

     0    0.5000    0.5000
0.5000    0.5000         0

theta =

1.5000
3.5000

The results seems quite right. But in the source code(line 94 ):
x[im+j] = (c[im+j] > theta)?(c[im+j]-theta):((c[im+j]< -theta)?(c[i*m+j]+theta):0);
when theta=1.5
mu(1,:) should be -0.5 0.5 0.5 in stead of 0 0.5000 0.5000

Please let me know, why is this?

Regards,
Hank

Currently we use FISTA, and however in practice, SPARSA works better. We may want to change to other solvers. Or, provide a list of solvers and ask users to choose one of them.

Hello, I'm using the sparse graph regularization to learn a model W of size 300 x #tasks. However, I only got 280 rows instead of 300. I'm guessing that some function removes rows of NaN in some step. Please confirm my conjecture. And is there a way to track which rows are removed? Thanks.

I want to use multi-task module to implement image classification. However, after reading your manual, I find that your data is not images. So I want to know can I use these modules in image classification? How can I do it? @jiayuzhou
Thanks!

We need to change the parameter to key/value, and users can leave some parameters blank to use default values.

We need to unify the input/output format.

in the least lasso demo，waht resolution is used？ is there a paper so that i can know more detailed information about the solution of lasso problem?

How to perform cross validation for Least SRMTL, which has two parameters? Thank you.