CODEWITHSUNDEEP
tensorflowmachine-learningdeep-learningkeras
J4m3s
J4m3s

Reputation: 43

What is the best way to save a keras model to resume training in the future?

I've searched about this a lot. In keras documentation it says that model.save() saves all the relevant info, i.e. model architecture, weights, optimizer state,...

Some other posts here on stackoverflow, mentioned saving the weights and loading them in the future to resume training, but the answers said that it's wrong, because it's not saving the optimizer state. I use callbacks to save best model base om validation accuracy, and it only save weights.

If weights aren't enough to resume training, why does call backs only save weights? Just for evaluating on the test set?

How can I properly save the best model then? Why doesn't call backs use model.save() to store all the info? How can i achieve this?

Upvotes: 1

Views: 861

Answers (2)

Keunwoo Choi
Keunwoo Choi

Reputation: 904

Definitely model.save() for all the reasons you've already mentioned. In a ModelCheckPoint callback, save_weights_only=false would do it. Set save_best_only=True if you wanna save some space or avoid clutters. This is equivalent to model.save().

If weights aren't enough to resume training, why does call backs only save weights? Just for evaluating on the test set?

Yes.

Upvotes: 2

Mohamed Elzarei
Mohamed Elzarei

Reputation: 546

According to Keras ModelCheckPointCallback documentation it saves either the whole model or only the weights according to save_weights_only flag which is False by default.

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

Upvotes: 2

Related Questions