During training, how to get value of a tensor defined during building time?

From this tutorial, one sees that the loss_value variable given as param to tape.gradient() can have a value like so:

tf.Tensor(0.36069894, shape=(), dtype=float32)

I am constructing something similar (code is very long and complex, not possible to put it here), but the loss element I access to is computed during building time (so, before the training loop), in this fashion:

x = Input(shape=(2,*input_shape))
z = SomeLayer(x)
y = SomeOtherLayer(z)
total_loss = TensorOperations(z)
net = Model(x,y)

Where TensorOperations can be different operations such as K.mean, K.abs etc.

But when I print out this total_loss variable during training:

with tf.GradientTape() as tape:
    net(input_data)
    print(total_loss)

I get something like:

Tensor("truediv_8:0", shape=(?,), dtype=float32)

My question is: how do I get a tensor with an actual value (like the first tensor) ?

[EDIT] the "real" code I am referring to is much more complex, here is the total_loss variable and here is the net = Model()

The easiest is probably to include your loss layer as an output of your model:

net = Model(x,[y, total_loss])

And then, during the training loop:

with tf.GradientTape() as tape:
    y_pred, loss_value = net(input_data)
    print(loss_value)

However, I would advise to try to decouple the loss from the model declaration, it's just more flexible.