I'm using classical PCA from scikit-learn for a classification model using LogisticRegression for a dataframe having 22 000 rows * 45 000 columns, data are scaled, but both have convergence issues when I upgraded python 3.6 to 3.7 and scikit-learn 0.19 to 0.23.
raise LinAlgError("SVD did not converge") LinAlgError: SVD did not converge
I have no NaN in my dataframe, but it can be a memory issue.
from sklearn.decomposition import PCA
n_components = int(X_main.shape[0] / 5)
pca = PCA(n_components=n_components)
principalComponents = pca.fit_transform(X_main)
principalDf = pd.DataFrame(data=principalComponents)
I changed nothing else, I resolved this by changing some hyper-parameters but i wonder why this appears, I've tested when I rollback to older versions and convergence error "disappear".
I don't know why, is there a warning now for convergence or is it computing another way, or other reasons ?
Thanks for you advice :)