Reproducing an issue with predict_proba() returning NaN values for linear classifiers with log loss in scikit learn

predict_proba() NaN issue

import sklearn
import sklearn.metrics
from sklearn.datasets import make_classification
from sklearn.linear_model import SGDClassifier
import numpy as np

# Create some training and testing data
data_size = 10000
train_size = 9000
test_size = data_size - train_size
X, y = make_classification(n_samples=data_size, 
                            n_features=500, 
                            n_informative=500, 
                            n_redundant=0, 
                            n_repeated=0, 
                            n_classes=10, 
                            n_clusters_per_class=1, 
                            weights=None, 
                            flip_y=0, 
                            class_sep=1, 
                            hypercube=True, 
                            shift=0.0, 
                            scale=1.0, 
                            shuffle=True, 
                            random_state=42)
X_test = X[train_size:,:]
y_test = y[train_size:]
X = X[:train_size]
y = y[:train_size]

# Standardize the data
X_mean = np.mean(X,axis=0)
X_std = np.std(X,axis=0)
X = (X - X_mean) / X_std
X_test = (X_test - X_mean) / X_std

# Create an SGDClassifier using log loss, default parameters
model = SGDClassifier(loss="log")

# Do a partial fit on a minibatch of 50 samples
# NB: if you try with the data generated above, but train on e.g. the first 42 samples,
# then no NaNs are returned...
model.partial_fit(X[0:50,:], y[0:50], classes=np.arange(10))

# Predict the class probabilities for all of the test data
y_prob = model.predict_proba(X_test)

# Print whether any of the probabilities are NaN
print(np.where(np.isnan(y_prob)))

# This creates a downstream error when calculating the log_loss with sklearn.metrics.log_loss:
loss = sklearn.metrics.log_loss(y_test, y_prob)
print(loss)