ValueError: Array contains NaN or infinity in _assert_all_finite during LinearSVC

问题

I was trying to classify the wine data set here -http://archive.ics.uci.edu/ml/datasets/Wine+Quality using logistic regression (with method ='bfgs' and l1 norm) and caught a singular value matrix error(raise LinAlgError('Singular matrix'), in-spite of full rank [which I tested using np.linalg.matrix_rank(data[train_cols].values) ] .

This is how I came to the conclusion that some features might be linear combinations of others. Towards this, I experimented of using Grid search/LinearSVC - and I get the error below, along with my code & data-set .

I can see that only 6/7 features are actually "independent" - which I interpret when comparing the rows of x_train_new[0] and x_train (so I can get which columns are redundant)

    # Train & test DATA CREATION
    from sklearn.svm import LinearSVC
    import numpy, random
    import pandas as pd
    df = pd.read_csv("https://github.com/ekta1007/Predicting_wine_quality/blob/master/wine_red_dataset.csv")
#,skiprows=0, sep=',')


    df=df.dropna(axis=1,how='any') # also tried how='all' - still get NaN errors as below
    header=list(df.columns.values) # or df.columns
    X = df[df.columns - [header[-1]]] # header[-1] = ['quality'] - this is to make the code genric enough
    Y = df[header[-1]] # df['quality']
    rows = random.sample(df.index, int(len(df)*0.7)) # indexing the rows that will be picked in the train set
    x_train, y_train = X.ix[rows],Y.ix[rows] # Fetching the data frame using indexes
    x_test,y_test  = X.drop(rows),Y.drop(rows)


# Training the classifier using C-Support Vector Classification.
clf = LinearSVC(C=0.01, penalty="l1", dual=False) #,tol=0.0001,fit_intercept=True, intercept_scaling=1)
clf.fit(x_train, y_train)
x_train_new = clf.fit_transform(x_train, y_train)
#print x_train_new #works
clf.predict(x_test) # does NOT work and gives NaN errors for some x_tests


clf.score(x_test, y_test) # Does NOT work
clf.coef_ # Works, but I am not sure, if this is OK, given huge NaN's - or does the coef's get impacted ?

clf.predict(x_train)
552   NaN
209   NaN
427   NaN
288   NaN
175   NaN
427   NaN
748     7
552   NaN
429   NaN
[... and MORE]
Name: quality, Length: 1119

clf.predict(x_test)
76    NaN
287   NaN
420     7
812   NaN
443     7
420     7
430   NaN
373     5
624     5
[..and More]
Name: quality, Length: 480

The strange thing is that when I run clf.predict(x_train) I still see some NaN's - What am I doing wrong ?After all the model was trained using this, and this should NOT occur , right ?

According to this thread, I also checked that there are no null's in my csv file (though I relabeled the "quality' to 5 and 7 labels only (from range(3,10) How to fix "NaN or infinity" issue for sparse matrix in python?

Also - here's the data type of x_test & y_test/train...

x_test
<class 'pandas.core.frame.DataFrame'>
Int64Index: 480 entries, 1 to 1596
Data columns:
alcohol                 480  non-null values
chlorides               480  non-null values
citric acid             480  non-null values
density                 480  non-null values
fixed acidity           480  non-null values
free sulfur dioxide     480  non-null values
pH                      480  non-null values
residual sugar          480  non-null values
sulphates               480  non-null values
total sulfur dioxide    480  non-null values
volatile acidity        480  non-null values
dtypes: float64(11)

y_test
1     5
10    5
18    5
21    5
30    5
31    7
36    7
40    5
50    5
52    7
53    5
55    5
57    5
60    5
61    5
[..And MORE]
Name: quality, Length: 480

and Finally..

clf.score(x_test, y_test)

Traceback (most recent call last):
  File "<pyshell#31>", line 1, in <module>
    clf.score(x_test, y_test)
  File "C:\Python27\lib\site-packages\sklearn\base.py", line 279, in score
    return accuracy_score(y, self.predict(X))
  File "C:\Python27\lib\site-packages\sklearn\metrics\metrics.py", line 742, in accuracy_score
    y_true, y_pred = check_arrays(y_true, y_pred)
  File "C:\Python27\Lib\site-packages\sklearn\utils\validation.py", line 215, in check_arrays
  File "C:\Python27\Lib\site-packages\sklearn\utils\validation.py", line 18, in _assert_all_finite
ValueError: Array contains NaN or infinity.


#I also explicitly checked for NaN's as here -:
for i in df.columns:
    df[i].isnull()

Tip : Please also mention if my thought process on using LinearSVC is correct, given my use case, or should I use Grid-search ?

Disclaimer : Parts of this code have been built on suggestions in similar contexts from StackOverflow and miscellaneous sources - My real use case is just trying to access if this method is a good fit for my scenario. That's all.

回答1:

This worked. The only I had to really change was use x_test*.values* along with the rest of pandas Dataframes(x_train, y_train, y_test) . As pointed out the only reason was incompatibility between pandas df and scikit-learn(which uses numpy arrays)

 #changing your Pandas Dataframe elegantly to work with scikit-learn by transformation to  numpy arrays
>>> type(x_test)
<class 'pandas.core.frame.DataFrame'>
>>> type(x_test.values)
<type 'numpy.ndarray'>

This hack comes from this post http://python.dzone.com/articles/python-making-scikit-learn-and and @AndreasMueller - who pointed out the inconsistency.

来源：https://stackoverflow.com/questions/21390084/valueerror-array-contains-nan-or-infinity-in-assert-all-finite-during-linearsv

标签

python-2.7

pandas

scikit-learn

svc