optimal loss function

[rrichards7]

What is the best loss function for multi-class accuracy evaluation metric?

The following should be explored:

1. Categorical binary cross entropy
2. Focal Loss (https://arxiv.org/pdf/1708.02002.pdf, https://www.kaggle.com/iafoss/pretrained-resnet34-with-rgby-0-460-public-lb)
3. AUC-PR Loss
4. AUC-ROC Loss
5. F1 Loss (https://www.kaggle.com/rejpalcz/cnn-128x128x4-keras-from-scratch-lb-0-328)

[rrichards7]

F1 loss:

def f1_loss(y_true, y_pred):
    
    #y_pred = K.cast(K.greater(K.clip(y_pred, 0, 1), THRESHOLD), K.floatx())
    tp = K.sum(K.cast(y_true*y_pred, 'float'), axis=0)
    tn = K.sum(K.cast((1-y_true)*(1-y_pred), 'float'), axis=0)
    fp = K.sum(K.cast((1-y_true)*y_pred, 'float'), axis=0)
    fn = K.sum(K.cast(y_true*(1-y_pred), 'float'), axis=0)

    p = tp / (tp + fp + K.epsilon())
    r = tp / (tp + fn + K.epsilon())

    f1 = 2*p*r / (p+r+K.epsilon())
    f1 = tf.where(tf.is_nan(f1), tf.zeros_like(f1), f1)
    return 1-K.mean(f1)

[rrichards7]

Focal Loss:

gamma = 2.0
epsilon = K.epsilon()
def focal_loss(y_true, y_pred):
    pt = y_pred * y_true + (1-y_pred) * (1-y_true)
    pt = K.clip(pt, epsilon, 1-epsilon)
    CE = -K.log(pt)
    FL = K.pow(1-pt, gamma) * CE
    loss = K.sum(FL, axis=1)
    return loss