Non-negative coefficients with lars

[anoopamagarwal]

I am running this piece of code

import numpy as np
from sklearn.linear_model import Lars
import sklearn

flag = True
print sklearn.__version__
while flag:
    X = np.random.randn(10,5)
    Y = np.random.randn(10)
    l=Lars(fit_intercept=False, positive=True)

    l.fit(X,Y)
    print l.coef_
    for i in l.coef_ :
        if i < 0 :
            flag = False

I am getting the following output

0.19.0
[ 0.00000000e+00 0.00000000e+00 5.21722148e-01 3.10743837e-01
4.18778571e-04]
[ 0. 0. 0. 0.45541234 0.09542552]
[ 0.80960144 0.35327211 0.17399586 0.16972765 -0.02813728]

with negative coefficients

[jnothman]

thanks for the report. please make this do np.random.seed(i) for iteration i, and report which i it breaks on so that the issue is easily replicated.

[anoopamagarwal]

Code snippet

import numpy as np
from sklearn.linear_model import Lars
import sklearn

flag = True
print sklearn.__version__
s = 100
while flag:
    np.random.seed(s)
    X = np.random.randn(10,5)
    Y = np.random.randn(10)
    l=Lars(fit_intercept=False, positive=True)

    l.fit(X,Y)
    print l.coef_
    for i in l.coef_ :
        if i < 0 :
            flag = False
    s += 1

Output

0.19.0
[ 0.18760379  0.25412609  0.          0.          0.        ]
[ 0.          0.07212829  0.19939359  0.          0.0600625 ]
[ 0.          0.09812548  0.30880926  0.          0.21128258]
[ 0.          0.          0.          0.          0.25207725]
[ 0.21952703  0.07202002  0.44614742  0.          0.        ]
[ 0.03364774  0.          0.04648766  0.          0.        ]
[  2.66874987e-04   7.74587735e-01   0.00000000e+00   5.53611627e-01
   8.67075723e-01]
[ 0.          0.22650594  0.05970705  0.          0.24344585]
[ 0.06457963  0.          0.          0.03992725  0.        ]
[ 0.         -0.02069482  0.          0.47615768  1.25238536]

[jnothman]

And what is s at the end?

[anoopamagarwal]

@jnothman random seed incremented at every iteration

[anoopamagarwal]

@jnothman Sorry didn't understand your question earlier. s=109 at the breaking point

Code snippet

import numpy as np
from sklearn.linear_model import Lars
import sklearn

flag = True
print sklearn.__version__
s = 100
while flag:
    np.random.seed(s)
    X = np.random.randn(10,5)
    Y = np.random.randn(10)
    l=Lars(fit_intercept=False, positive=True)

    l.fit(X,Y)
    print l.coef_
    for i in l.coef_ :
        if i < 0 :
            print s
            flag = False
    s += 1

Output

0.19.0
[ 0.18760379  0.25412609  0.          0.          0.        ]
[ 0.          0.07212829  0.19939359  0.          0.0600625 ]
[ 0.          0.09812548  0.30880926  0.          0.21128258]
[ 0.          0.          0.          0.          0.25207725]
[ 0.21952703  0.07202002  0.44614742  0.          0.        ]
[ 0.03364774  0.          0.04648766  0.          0.        ]
[  2.66874987e-04   7.74587735e-01   0.00000000e+00   5.53611627e-01
   8.67075723e-01]
[ 0.          0.22650594  0.05970705  0.          0.24344585]
[ 0.06457963  0.          0.          0.03992725  0.        ]
[ 0.         -0.02069482  0.          0.47615768  1.25238536]
109

[jnothman]

So the minimal reproducing example is:

import numpy as np
from sklearn.linear_model import Lars
np.random.seed(109)
X = np.random.randn(10,5)
Y = np.random.randn(10)
l = Lars(fit_intercept=False, positive=True)
l.fit(X,Y)
print(l.coef_)
assert(np.all(l.coef_ >= 0))

Thanks. I can reproduce, indeed.

[jnothman]

And the bug appears to have been present since positive was introduced.

[VinodKumarLogan]

Hi @jnothman, Can I work on this issue?

[jnothman]

If you're confident you can debug it, you're more than welcome to!

[agramfort]

see #10248