Implement two functions.
smooth_labels(y_true, num_classes, epsilon) returns an array of shape (N, K) with K = num_classes. For a sample with label the target row is for the true class and for every other class (rows sum to 1). Raise ValueError if epsilon is outside .
label_smoothing_cross_entropy(logits, y_true, num_classes, epsilon=0.1, round_decimals=None) takes logits of shape (N, K) and integer labels y_true of shape (N,), and returns the mean over the samples of
computed with a numerically stable log-softmax (subtract the row maximum before exponentiating). Return a Python float; if round_decimals is an integer, round the loss to that many decimals first. Raise ValueError if logits is not 2-D or its second dimension differs from num_classes.