How to save epochs progress to log file?

[Lior Magen]

I'm using Keras for classification task and now want to deploy it in production. Since now when I was training I simply used verbose=1 or other value but I don't want the progress to be printed, I want it to be saved as logs. I'm using 'logging' library on Python and want the progress to be saved there.

I believe that I should use some Callback and tried using:

self.classifier.fit(x_tr, y_tr, batch_size=self.batch_size, nb_epoch=self.nb_epoch,
                    verbose=self.print_status, validation_data=(x_te, y_te), callbacks=[keras.callbacks.ProgbarLogger()])

It doesn't change anything and I was looking at Keras documentation but still no clue. 

Can anyone help me with how to use this callback properly please?

[jeffer...@gmail.com]

You can remove the `calback` argument altogether because the `History` callback "is applied automatically to every Keras model" (per https://keras.io/callbacks/). What I've done is just something like this:

import cPickle as pickle

# ...

results = model.fit(...) # results.history is dict of loss, acc, val_loss, and val_acc

# ...
with open('results.pickle', 'wb') as f:
    pickle.dump(results.history, f)

[Lior Magen]

This will keep the results after the training is done, right? I want something to print the progress in real time. Any optional Callback for this task maybe?

[Benjamin Striner]

Hi Lior,

Here is the callback that I use. Haven't PRed it in to keras yet. Just use model.fit(...,callbacks=[LoggingCallback(logging.info)])

Cheers,

Ben

class LoggingCallback(Callback):
    """Callback that logs message at end of epoch.
    """

    def __init__(self, print_fcn=print):
        Callback.__init__(self)
        self.print_fcn = print_fcn

    def on_epoch_end(self, epoch, logs={}):
        msg = "Epoch: %i, %s" % (epoch, ", ".join("%s: %f" % (k, v) for k, v in logs.iteritems()))
        self.print_fcn(msg)

[Benjamin Striner]

BTW, for saving results after training, I like to use pandas. Write all of the training history straight to CSV in a one-liner.

history= model.fit(...)

pandas.DataFrame(history.history).to_csv("history.csv")

[Benjamin Striner]

Just made a PR with the callback.

model.fit(...,callbacks=[LoggingCallback(logging.info)])

If you want to pull it in locally for now:

git remote add bstriner https://github.com/bstriner/keras.git

git fetch bstriner

git merge bstriner/logging_callback

https://github.com/fchollet/keras/pull/5059

[Klemen Grm]

You can also use the built-in CSVLogger callback.

https://keras.io/callbacks/#csvlogger

[jeffer...@gmail.com]

Yes, `keras.callback.CSVLogger` is the way to go. I hadn't stumbled across that until after my original reply. :-)

[Lior Magen]

Thank you so much for your help, this is exactly what I needed.

In the end Iv'e used your code with minor changes (really minor, just so it will work with Python 3.*)

class LoggingCallback(Callback):
    """Callback that logs message at end of epoch.
    """


    def __init__(self, print_fcn=print):
        Callback.__init__(self)
        self.print_fcn = print_fcn


    def on_epoch_end(self, epoch, logs={}):

        msg = "{Epoch: %i} %s" % (epoch, ", ".join("%s: %f" % (k, v) for k, v in logs.items()))
        self.print_fcn(msg)

Again, I really appreciate your help.