Trying to understand the difference between categorical_accuracy and sparse_categorical_accuracy in metrics.py

[travel...@gmail.com]

When I look at some of the metrics in metrics.py, two of the metrics I see are:

def categorical_accuracy(y_true, y_pred):
    return K.mean(K.equal(K.argmax(y_true, axis=-1),
                  K.argmax(y_pred, axis=-1)))




def sparse_categorical_accuracy(y_true, y_pred):
    return K.mean(K.equal(K.max(y_true, axis=-1),
                          K.cast(K.argmax(y_pred, axis=-1), K.floatx())))

Am I correct in believing that both of these metrics are designed for a 1-hot encoding, but that the sparse_categorical_accuracy casts outputs to type float, whereas the categorical_accuracy does not?

[boll...@gmx.de]

I believe that one is for use with the "categorical_crossentropy" loss, the other for use with the "sparse_categorical_crossentropy" loss. The difference should just be in the internal representation of the training data; with categorical crossentropy, you have 1-hot vectors, while with the sparse version you just use numeric indices.

- Marcel