I would have to agree. Accuracy is the better, fairer most sensible metric.

The objective should be to classify an image to a class. With log loss, we have to pay attention to the other (wrong) probabilities, which depend on the model, training, loss function used etc., so that they march whatever it is the marking scheme is.

If I predict the correct class e.g

0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0 ...

my log loss will be higher if the correct answer is

0.5, 0.6, 0.3, 0.4, 0.2, 0.9, 0.1, 0.2, 0.2

much as the argmax for both the logits is the same i.e. index 5

@zindi please reconsider. It is not too late