ENH add Lasso.jl julia solver

[jolars]

This PR adds support for the Lasso.jl Julia package. One problem with the solver is that it has no setting for maximum number of iterations, so when tolerance is low it may fail to converge. I've filed an issue at the repo about this and plan to file a PR eventually to get this fixed.

[tomMoral]

Overall, this looks super nice!! Thx @jolars

A few comments along the way.

[tomMoral]

can you comment on what this is?

[jolars]

randomize shuffles indices before cd loops (sometimes or always, I don't know), stopearly is for stopping the path (change in deviance, saturation), maxncoef is also related to stopping the path early (maximum number of nonzero coefficients). I ran into a bug in the package when these were at their defaults, but I think only the maxncoef setting is actually necessary to work around it. I'll experiment a bit. randomize should really be true I think, but the others do not really matter.

[tomMoral]

For the broken test, we need to skip the julia solver for OSX as this causes a segfault in this case.

[tomMoral]

LGTM! thx @jolars

[agramfort]

@mathurinm merge if happy

[mathurinm]

It's running fine on simulated on my machine.:

I only get a lot of

/home/mathurin/anaconda3/envs/benchopt_lasso/lib/python3.9/site-packages/julia/core.py:703: FutureWarning: Accessing `Julia().<name>` to obtain Julia objects is deprecated.  Use `from julia import Main; Main.<name>` or `jl = Julia(); jl.eval('<name>')`.
  warnings.warn(

can we fix it of silence it ? sorry if this has been discussed already

[jt2gtwci]

I only get a lot of

/home/mathurin/anaconda3/envs/benchopt_lasso/lib/python3.9/site-packages/julia/core.py:703: FutureWarning: Accessing `Julia().<name>` to obtain Julia objects is deprecated.  Use `from julia import Main; Main.<name>` or `jl = Julia(); jl.eval('<name>')`.
  warnings.warn(

can we fix it of silence it ? sorry if this has been discussed already

We talked about this a bit in benchopt/benchopt#372. I actually think the warning might be a bug: JuliaPy/pyjulia#497

[mathurinm]

So we silence the warning in the solver for now ?

[jolars]

my setup is segfaulting for some reason right now, but I submitted a patch to ignore the warnings somewhat blindly. Can anybody please check that it works?

[mathurinm]

can you send me the command that causes the segfault ?

[jolars]

can you send me the command that causes the segfault ?

I just tried to call benchopt run -e . -s lasso_jl -d simulated, but I'm sure it's not related to benchopt in any way, just some annoying result of version/compilation problems from the pyenv, conda, and julia mix.

[mathurinm]

It works but:

• it takes around 1 min for the solver to start
• I still get the warnings
• for some runs it outputs error

[jolars]

it takes around 1 min for the solver to start

that sounds like JIT compiliation

I still get the warnings

Hm, strange. I can silence the same warning in a small script just by doing the same thing:

from julia import Julia
import warnings

warnings.filterwarnings("ignore", category=FutureWarning)
jl = Julia(compiled_modules=False)
jl.eval("1 + 1")

for some runs it outputs error

What kind of errors? I used to receive errors previously due to convergence issues but I thought I had fixed them with a try catch statement.

Are you sure that you don't have results cached or something?

[mathurinm]

There seems to be something wrong with sparse datasets:

(benchopt_lasso) ➜  lasso git:(bench) ✗ benchopt run . -s lasso_jl -d "libsvm[rcv1.binary]" -r1
WARNING: astropy not found, will default to scipy for convolution
Benchopt is running
libsvm[dataset=rcv1.binary]                                                                      
  |--Lasso Regression[fit_intercept=True,reg=0.5]                                                
/home/mathurin/anaconda3/envs/benchopt_lasso/lib/python3.9/site-packages/julia/core.py:703: FutureWarning: Accessing `Julia().<name>` to obtain Julia objects is deprecated.  Use `from julia import Main; Main.<name>` or `jl = Julia(); jl.eval('<name>')`.
  warnings.warn(
  |--Lasso Regression[fit_intercept=True,reg=0.1]                                                
  |--Lasso Regression[fit_intercept=True,reg=0.05]                                               
  |--Lasso Regression[fit_intercept=True,reg=0.01]                                               
  |--Lasso Regression[fit_intercept=True,reg=0.001]                                              
  |--Lasso Regression[fit_intercept=False,reg=0.5]                                               
  |--Lasso Regression[fit_intercept=False,reg=0.1]                                               
  |--Lasso Regression[fit_intercept=False,reg=0.05]                                              
  |--Lasso Regression[fit_intercept=False,reg=0.01]                                              
  |--Lasso Regression[fit_intercept=False,reg=0.001]                                             
No output produced.                                                                              
Saving result in: None                                                                           

It's because sparse datasets are skipped by benchopt for julia, I remove this in
benchopt/benchopt#404

[mathurinm]

I don't get it to converge on rcv1.binary, for fit_intercept=False or True:

Same behavior for smaller reg (0.01, 0.001)

[jolars]

I am getting the following error when trying to run this now:

signal (11): Segmentation fault
in expression starting at /home/gerd-jln/.conda/envs/benchopt_benchmark_lasso/lib/python3.8/site-packages
/julia/install.jl:34
PyVectorcall_Function at /tmp/python-build.20220215103121.191136/Python-3.9.10/./Include/cpython/abstract
.h:73 [inlined]
_PyObject_Call at /tmp/python-build.20220215103121.191136/Python-3.9.10/Objects/call.c:265
unknown function (ip: 0x7ffcf9909daf)
Allocations: 5608687 (Pool: 5606149; Big: 2538); GC: 6

I am guessing that this has something to do with the mismatch in versions between conda and whatever this other python 3.9.10 version is doing there

[jolars]

I don't get it to converge on rcv1.binary, for fit_intercept=False or True:

Same behavior for smaller reg (0.01, 0.001)

Hm, could it be that the iteration limit is not sufficiently large?

[mathurinm]

Ah, because it behaves like glmnet and returns a vector of 0s if it does not converge for the desired lambda ?