CODEWITHSUNDEEP
pythontensorflowkerascallback
John Sarle
John Sarle

Reputation: 31

Is it possible to log the epoch results in the tf.keras model using a tensorflow callback, in order to save at the end of training?

I am trying to save the logs value that gets passed into a tf.keras callback at the end of every epoch to track how the model did along the way.

I have tried writing a custom tf.keras callback that appends the logs value that is passed to the callback to an array that I initialized at the beginning of training using the same callback. However upon debugging after the first epoch the array that I initialized in the model is None at the end of the second epoch. Below is the custom callback I created. 

class LogEpochScores(tf.keras.callbacks.Callback):
    def __init__(self):
        super(LogEpochScores, self).__init__()

    def on_train_begin(self, logs=None):
        self.model.epoch_log = []

    def on_epoch_end(self, epoch, logs=None):
        self.model.epoch_log = self.model.epoch_log.append(logs)

I expected the following array at the end of training 

[
{loss: 1, acc:1, val_loss:1, val_acc:1},
{loss: 2, acc:2, val_loss:2, val_acc:2},
{loss: 3, acc:1, val_loss:3, val_acc:3},
{loss: 4, acc:1, val_loss:4, val_acc:4},
{loss: 5, acc:1, val_loss:5, val_acc:5},
{loss: 6, acc:1, val_loss:6, val_acc:6}
]

Each entry being the results from the respective epoch.

The output actually is None

Edit: Formatting

Upvotes: 1

Views: 1683

Answers (1)

John Sarle
John Sarle

Reputation: 31

Apparently I am not intelligent. I was using append the wrong way. Its destructive so you don't set it as a variable after use. The corrected code is as follows:

class LogEpochScores(tf.keras.callbacks.Callback):
    def __init__(self):
        super(LogEpochScores, self).__init__()

    def on_train_begin(self, logs=None):
        self.model.epoch_log = []

    def on_epoch_end(self, epoch, logs=None):
        self.model.epoch_log.append(logs)

Notice in on_epoch_end I change 

self.model.epoch_log = self.model.epoch_log.append(logs)

to 

self.model.epoch_log.append(logs)

Additionally, in the tf.kera model object, there is an attribute called history which is exactly what I needed. So in the end there was no need for this callback

Upvotes: 2

Related Questions