What exactly is Keras loss in history
Verfasst: Samstag 19. Dezember 2020, 13:22
Hello AI-Friends,
I am using a callback function to calculate train and test error after each epoch end by calling model.evaluate() . However if I compare my calculated train loss from model.evaluate(x_train, y_train) it is different from the loss saved in history.history['loss']. The calculated test loss from my callback is identical to the history.history['val_loss'].
So I wonder how does Keras calculate the train loss and what exactly is saved in history.history['loss'] ?
I have to calculate the loss after each epoch for different datasets, since I want to evaluate the performance of losses for a sequential training of multiple datasets.
Anybody have an idea why these losses for training data are not identical. Is there a better way to do it?
This is my code:
I am using a callback function to calculate train and test error after each epoch end by calling model.evaluate() . However if I compare my calculated train loss from model.evaluate(x_train, y_train) it is different from the loss saved in history.history['loss']. The calculated test loss from my callback is identical to the history.history['val_loss'].
So I wonder how does Keras calculate the train loss and what exactly is saved in history.history['loss'] ?
I have to calculate the loss after each epoch for different datasets, since I want to evaluate the performance of losses for a sequential training of multiple datasets.
Anybody have an idea why these losses for training data are not identical. Is there a better way to do it?
This is my code:
Code: Alles auswählen
class MyCustomCallback(keras.callbacks.Callback):
def __init__(self):
self.results = {
'eval_train' : {},
'eval_test' : {}
}
def on_epoch_end(self, epoch, logs=None):
eval_train = self.model.evaluatex_train, y_train, verbose=1)
eval_test = self.model.evaluate(x_test, y_test, verbose=1)
self.results['eval_train'].append(eval_train)
self.results]['eval_test'].append(eval_test)
myCallback = MyCustomCallback()
history = model.fit(x_train, y_train,
epochs=10, batch_size=256, verbose=1,
callbacks=[myCallback]),validation_data=(x_test, y_test)