Why does model.losses return regularization losses?

Question

I have met a snippet of code of tensorflow 2.0, which is used for calculating the loss. The total loss is composed of two parts: 1) regularization loss, 2) prediction loss. My question is why model.losses is regularization loss? model here is an instance of tf.keras.Model. I'm kind of confused by the tensorflow official API documentation. tf.keras.Model, it says

Losses which are associated with this Layer.

Variable regularization tensors are created when this property is accessed, so it is eager safe: accessing losses under a tf.GradientTape will propagate gradients back to the corresponding variables.

Why could we get regularization loss via accessing losses property? Also, what is eager safe? If losses property is returning regularization loss, why is it named losses instead of regularization_loss?

with tf.GradientTape() as tape:
  outputs = model(images, training=True)
  regularization_loss = tf.reduce_sum(model.losses)
  pred_loss = ...
  total_loss = pred_loss + regularization_loss

Oocyte

Answer 1

We get regularization losses accessing the losses property because these losses are created during the model definition. Since the model is a Keras model, you've built it using Keras layers. Every Keras layer (Dense, Conv3D, ...) can be regularized and this is a property of the layer itself.

The model, being an ordered collection of layers, contains all the layers losses inside the losses property.

Eager safe means you can use the losses property of the model during the eager training, being sure that the gradient is propagated only to the correct layers. E.g. if you added an l2 regularization only on the second layer of a model, the variables of the second layer only are influenced (and updated) by that term of the loss.

Is named losses instead of regularization_losses because is not limited to regularization losses only; when you compile a model, a non-regularization loss is added to that property

Contraposition