jasonge27 / picasso Goto Github PK

1. in the python interface, size_act is set to size of (nlambda, d), however, it should be set to size (nlambda).
2. in c code actgd.cpp, model_intercept is updated whatever use_intercept is set to true or false!

I was wondering what link scale is assumed for Poisson models? Is this a log link? If so, I was wondering if it would be hard to also support Poisson models with an identity link? I am asking because I am dealing with models where my noise is Poisson but where my covariates act additively rather than multiplicatively...

In

Shouldn't it be like this:
`double soft_thresh_mcp(double y, double lamb, double gamma){

if(fabs(y)>fabs(gamma*lamb)) {
    return y;
}else{
    return soft_thresh_l1(y, lamb)/(1-1/gamma);
}

}`

Would it be possible for picasso in the future to also support box constraints on the fitted coefficients by any chance, similar to what glmnet can do using arguments lower.limits and upper.limits? E.g. to support nonnegativity constraints?

HI,

First off, thank you for putting this together. It has helped me understand SCAD and play around with the model.

I noted that after fitting the model using the Python module, running predict multiple times will change the model.

For example

>>> scadpath.result['beta'][4]
[ 6.56925953e-02  1.01571611e+03 -3.36603532e-11  3.20955197e-03
  4.31731825e+03 -7.09569678e-03 -9.05913159e-01  2.02699282e-17]
>>> predictionPath = scadpath.predict(X_train, lambdidx=4)
>>> scadpath.result['beta'][4]
[ 5.28674587e-02  2.13116762e+03 -4.48672601e-12  2.19283383e-03
  1.02795980e+04 -5.06569449e-03 -1.27797957e+00  6.91764767e-19]

Please let me know if I am misinterpreting the results in any way.

Thank you,

predictionScad = scad.predict(X_train)

Thanks for such a fast package. I am using the python pycasso on a subset of big data for design matrix X. It works for X with dimensions 210,000 x 123,916 , but it fails with Segmentation fault for X with dimensions 235,495 x 123,916 (X is always a subset of the same data). Debugging with gdb shows

In the python the error is thrown in pycasso/core.py, line 209 in wrapper.
What do you suggest?

I've accidentally noticed that the default stopping precision differs across the files. The documentation states that the default value is 1e-7. At the same time the python package and some of the R routines set it to 1e-4. What should the default value be?

the intercept argument seems not working in python.

Great library. There is feature though that is stopping me from using it and that is support for something like the penalty.factor in glmnet.

I'm wondering if there's any possiblity that picasso supports long vectors so that it works with big matrices of many columns and rows (.e.g 22000000 rows, 120 columns, more elements than max integer).

It is unfortunate that glmnet does not support long vectors since it's not trivial to adapt its Fortran implementation to work with long vectors that uses double length.

Greetings.

According to python-package/pycasso/lib/core.py, the mapping of regularizer is as follows:

• MCP = 2, SCAD = 3, and 1 for others

However, in src/c_api/c_api.cpp, the mapping of regularizer is as follows:

• L1 = 1, MCP = 2, SCAP for all the others

This will cause incompatible situations. An exception should be raised for number not in 1~3, and mapping logic should be the same.

On the other hand, when it comes to ActGDSolver::solve(), regfunc is determined by enum defined in include/picasso/solver_params.hpp. However, according to the cpp standard, the default value of enum is set to 0, which means L1 = 0, SCAD =1 and MCP = 2, and this will cause wrong regfunc to be applied.

Please correct me if I was wrong, thank you.
Wenny

Hello, it is a nice work!
I found there is a spelling mistake in core.py line 130 which "lamdas.size" should be "lambdas.size". Would you please check this? Thanks!