In tensorflow 1.15.5 / keras, how to use a tensor defined during building as a loss value for tape.gradient

I want to change this line of a model training :

net.train_on_batch(train_subdata_batch_cache, train_subdata_to_batch_cache)

Into something like this:

with tf.GradientTape() as tape:
    net(tf.convert_to_tensor(train_subdata_batch_cache, dtype=tf.float32))
    loss_value = self.loss
grads = tape.gradient(loss_value, net.trainable_weights)
optimizer.apply_gradients(zip(grads, net.trainable_weights))

Where self.loss comes from here

But, when I execute this code, I get:

> AttributeError: 'function' object has no attribute 'dtype'

So I also tried to add a new variable in model.py (just below l. 1534), namely:

self.otherloss = total_loss

and then do:

loss_value = self.otherloss

But this time I am getting:

> AttributeError: 'RefVariable' object has no attribute '_id'

To sum up, I want to take the loss stored in the variable total_loss and use it in my custom training loop.

Thanks

To my understanding the problem is that net is the layer of a NN with its own loss cost, and self.loss is a function in the class Network working as an interface to execute logic of several layers.

I think what you are searching for the loss of a layer which it's in the object net, but I am not sure which object is that to get the loss, you can try to use the Keras api to check the loss by:

tf.keras.losses.get(net)

Or you can try to check Tf Keras keep loss of every batch

If instead you search for the list of the losses of all the layers you can find it in self.losses