Multiclass SVM failed to use 20 News Group Dataset

Question

I'm trying to use multiclass SVM code from Mblondel Multiclass SVM, I read his paper and he used dataset from sklearn 20newsgroup, but when I tried to use that, the code isn't working properly.

I tried to change the code to match 20newsgroup dataset. but I'm stuck at this error..

Traceback (most recent call last):

File "F:\env\chatbotstripped\CSSVM.py", line 157, in

clf.fit(X, y)

File "F:\env\chatbotstripped\CSSVM.py", line 106, in fit

v = self._violation(g, y, i)

File "F:\env\chatbotstripped\CSSVM.py", line 50, in _violation

elif k != y[i] and self.dual_coef_[k, i] >= 0:

IndexError: index 20 is out of bounds for axis 0 with size 20

this is the main code:

from sklearn.datasets import fetch_20newsgroups
news_train = fetch_20newsgroups(subset='train')
X, y = news_train.data[:100], news_train.target[:100]

clf = MulticlassSVM(C=0.1, tol=0.01, max_iter=100, random_state=0, verbose=1)
X = TfidfVectorizer().fit_transform(X)
clf.fit(X, y)
print(clf.score(X, y))

this is the fit code:

def fit(self, X, y):
    n_samples, n_features = X.shape

    self._label_encoder = LabelEncoder()
    y = self._label_encoder.fit_transform(y)

    n_classes = len(self._label_encoder.classes_)
    self.dual_coef_ = np.zeros((n_classes, n_samples), dtype=np.float64)
    self.coef_ = np.zeros((n_classes, n_features))

    norms = np.sqrt(np.sum(X.power(2), axis=1)) # i changed this code

    rs = check_random_state(self.random_state)
    ind = np.arange(n_samples)
    rs.shuffle(ind)

    # i added this sparse
    sparse = sp.isspmatrix(X)
    if sparse:
        X = np.asarray(X.data, dtype=np.float64, order='C')

    for it in range(self.max_iter):
        violation_sum = 0
        for ii in range(n_samples):
            i = ind[ii]
        
            if norms[i] == 0:
                continue
        
            g = self._partial_gradient(X, y, i)
            v = self._violation(g, y, i)
            violation_sum += v
         
            if v < 1e-12:
                continue

            delta = self._solve_subproblem(g, y, norms, i)
            self.coef_ += (delta * X[i][:, np.newaxis]).T
            self.dual_coef_[:, i] += delta

        if it == 0:
            violation_init = violation_sum

        vratio = violation_sum / violation_init

        if self.verbose >= 1:
            print("iter", it + 1, "violation", vratio)

        if vratio < self.tol:
            if self.verbose >= 1:
                print("Converged")
            break
    return self

and _violation code:

def _violation(self, g, y, i):
    smallest = np.inf
    for k in range(g.shape[0]):
        if k == y[i] and self.dual_coef_[k, i] >= self.C:
            continue
        elif k != y[i] and self.dual_coef_[k, i] >= 0:
            continue

        smallest = min(smallest, g[k].all()) # and i added .all()
    return g.max() - smallest

I know there's something wrong with the index, I'm not sure how to fix that, and I don't want to ruin the code because I don't really understand how's this code works.

Answer 1

You have to convert the sparse matrix output of tfidf vectorizer into dense matrix then make it as a 2D array. Try this!

from sklearn.datasets import fetch_20newsgroups
from sklearn.feature_extraction.text import TfidfVectorizer
news_train = fetch_20newsgroups(subset='train')
text, y = news_train.data[:1000], news_train.target[:1000]

clf = MulticlassSVM(C=0.1, tol=0.01, max_iter=100, random_state=0, verbose=1)
vectorizer= TfidfVectorizer(min_df=20,stop_words='english')
X = np.asarray(vectorizer.fit_transform(text).todense())
clf.fit(X, y)
print(clf.score(X, y))

Output:

iter 1 violation 1.0
iter 2 violation 0.07075102408683964
iter 3 violation 0.018288133735158228
iter 4 violation 0.009149083942255389
Converged
0.953