Why does LassoCV mean square error depend on alphas interval?

Question

I am tuning alpha coefficient in Lasso regularisation to get best result on cross-validation:

alphas = np.arange(1,200, 1)
from sklearn.linear_model import LassoCV
lasso=LassoCV(alphas=alphas)
r=lasso.fit(X,y)
plt.plot(alphas,lasso.mse_path_)

This gives me following picture:

I want to see minimum of the red curve,enlarge alpha interval twice (np.arange(1,400, 1)) and rerun the code:

Why curves' coorientations depend on alphas array size??

Answer 1

You should use sklearn.linear_model.LassoCV.alphas_ for the $X$ axis.

Your plot

plt.plot(alphas,lasso.mse_path_)

implicitly assumes the order is the original one of alphas. This is a logical assumption, but it happens to be incorrect - in your case, it's outputting results corresponding to a reverse order of your alphas. If you use

plt.plot(lasso.alphas_, lasso.mse_path_)

your plots will do what you want.