Python – Keras custom loss calculation is incorrect

I’m trying to use a custom loss function in Keras. My implementation looks like this:

class LossFunction:
    ...

def loss(self, y_true, y_pred):
        ...
        localization_loss = self._localization_loss()
        confidence_loss = self._object_confidence_loss()
        category_loss = self._category_loss()

self.loc_loss = localization_loss
        self.obj_conf_loss = confidence_loss
        self.category_loss = category_loss

tot_loss = localization_loss + confidence_loss + category_loss
        self.tot_loss = tot_loss
        return tot_loss

Then I define custom metrics to see stored tensors, for example:

class MetricContainer:
    def __init__(self, loss_obj):
        self.loss = loss_obj

def local_loss(self, y_true, y_pred):
        return self.loss.loc_loss

def confidence_loss(self, y_true, y_pred):
        return self.loss.obj_conf_loss

def category_loss(self, y_true, y_pred):
        return self.loss.category_loss

def tot_loss(self, y_true, y_pred):
        return self.loss.tot_loss

Then I compile my model with this command :

model.compile('adam', 
              loss=loss_obj.loss,
              metrics=[metric_container.local_loss, 
                       metric_container.confidence_loss, 
                       metric_container.category_loss, 
                       metric_container.tot_loss])

When I train the model (on a very small training set), I get the following output:

Epoch 1/2
1/2 [==============>...............] - ETA: 76s - loss: 482.6910 - category_loss: 28.1100 - confidence_loss: 439.9192 - local_loss: 13.1180 - tot_loss: 481.1472 
2/2 [==============================] - 96s - loss: 324.6292 - category_loss: 18.1967 - confidence_loss: 296.0593 - local_loss: 8.8204 - tot_loss: 323.0764 - val_loss: 408.1170 - val_ category_loss: 0.0000e+00 - val_confidence_loss: 400.0000 - val_local_loss: 6.5036 - val_tot_loss: 406.5036

For some reason, tot_loss and loss don’t match, even though I should be using the same value for them.

Any idea why this is the case? Will Keras make some modifications after returning?

Your loss is equal to the sum of the selected loss function and the regularization term. So, if you use any type of regularization – it affects your losses by adding regularization items.