Sklearn StratifiedKFold: ValueError: Supported target types are: ('binary', 'multiclass'). Got 'multilabel-indicator' instead

keras.utils.to_categorical produces a one-hot encoded class vector, i.e. the multilabel-indicator mentioned in the error message. StratifiedKFold is not designed to work with such input; from the split method docs:

split(X, y, groups=None)

[...]

y : array-like, shape (n_samples,)

The target variable for supervised learning problems. Stratification is done based on the y labels.

i.e. your y must be a 1-D array of your class labels.

Essentially, what you have to do is simply to invert the order of the operations: split first (using your intial y_train), and convert to_categorical afterwards.

Call to split() like this:

for i, (train_index, val_index) in enumerate(kf.split(x_train, y_train_categorical.argmax(1))):
    x_train_kf, x_val_kf = x_train[train_index], x_train[val_index]
    y_train_kf, y_val_kf = y_train[train_index], y_train[val_index]

I bumped into the same problem and found out that you can check the type of the target with this util function:

from sklearn.utils.multiclass import type_of_target
type_of_target(y)

'multilabel-indicator'

From its docstring:

• 'binary': y contains <= 2 discrete values and is 1d or a column vector.
• 'multiclass': y contains more than two discrete values, is not a sequence of sequences, and is 1d or a column vector.
• 'multiclass-multioutput': y is a 2d array that contains more than two discrete values, is not a sequence of sequences, and both dimensions are of size > 1.
• 'multilabel-indicator': y is a label indicator matrix, an array of two dimensions with at least two columns, and at most 2 unique values.

With LabelEncoder you can transform your classes into an 1d array of numbers (given your target labels are in an 1d array of categoricals/object):

from sklearn.preprocessing import LabelEncoder

label_encoder = LabelEncoder()
y = label_encoder.fit_transform(target_labels)