Question about loss function

[jonathanventura]

Why is the loss function

    loss = tf.reduce_sum(K.square(target - y_pred*mask)) / tf.reduce_sum(mask)

instead of

    loss = tf.reduce_sum(K.square(target - y_pred)*mask) / tf.reduce_sum(mask)

?

[tibuch]

Hi @jonathanventura

We initialize target with zeros and only set values at the active pixels. So target - y_pred*mask is zero everywhere except at the active pixels.

Your suggested loss should lead to the same result and looks like the safer choice to be honest :)

[jonathanventura]

Ah, that makes sense -- thank you!