get name of loss function used in keras model

Question

I want to save the name of the loss function I used in my keras model. I looked into the documentation but haven't found a way to get this name. If possible I also want to save this name in case I use a custom loss function. Or at least extract the information from the model that I've used a custom loss function. This is what my model is looking like:

model = Sequential()
model.add(Dense(5, input_dim=4, activation='tanh'))
model.add(Dense(5, activation='tanh'))
model.add(Dense(5, activation='tanh'))
model.add(Dense(3))
model.compile(loss='mean_squared_error', optimizer='nadam', metrics=['accuracy'])

And for custom loss:

model.compile(loss=custom_loss, optimizer='nadam', metrics=['accuracy'])

Answer 1

The loss is saved as an attribute inside the model object. I was not able to find it in the docs, I found it using dir(model) . You can retrieve from the attribute the name of the loss function, a tf.keras.losses.Loss instance or a custom callable.

model.compile(loss='mean_squared_error', optimizer='nadam', metrics=['accuracy'])
model.loss
>>> 'mean_squared_error'

model.compile(loss=tf.keras.losses.MeanSquaredError(), optimizer='nadam', metrics=['accuracy'])
model.loss
>>> <keras.losses.MeanSquaredError at 0x7f5d47ee4710>

def my_loss_fn(y_true, y_pred):
    squared_difference = tf.square(y_true - y_pred)
    return tf.reduce_mean(squared_difference, axis=-1)
model.compile(loss=my_loss_fn, optimizer='nadam', metrics=['accuracy'])
model.loss
>>> <function __main__.my_loss_fn>

Answer 2

you can still access the loss name as string in the last two cases of Santiago's answer:

• for tf.keras.losses objects: model.loss.name

• for custom loss functions: model.loss.__name__ (this will return the variable name used to define the function)