How to retrain/update keras model?

Question

How to update/append new data to my model without starting to retrain from scratch? My dataset are images and the output is to predict emotion.

model.fit(x=train_image, y=train_label, epochs=1, batch_size=1)

Model.fit seems not appending my new data but overwriting the model. My output is only one (the last one that I updated).

Loading a trained Keras model and continue training

I've already search for this but doesn't work as well.

Edit 1: What can we do when we lose our previously trained data. In simple words we lose our training data just after the training is completed and we can not get the data back again, just the thing we have to retain is the learning done from it and also retrain the model when we receive new data.

What exactly is not working? Do you get any error messages? Do you also save the state of your optimizer somewhere? — Hagbard, Jan 25 '19 at 10:25
@Hagbard Yes, I do save it and there's no error with it. The problem is it keeps on overwriting my model everytime I retrain/update it using model.fit. — lambda, Jan 27 '19 at 23:56

score 3 · Accepted Answer · answered Jan 25 '19 at 12:06

3

It is really easy. For instance model.save() in Keras can save the model weights, which is what you actually need to keep.

After that you can add new images to your training set, but just use the weights you saved.

Every new epoch will pass trough all the images (new and old).

answered Jan 25 '19 at 12:06

prosti

42,291
14
186
151

2

Well, I saved the x and y array of training data to numpy file. Then append the new x and y data to that file and later retrain the whole model . And it's more of a data gathering method. – lambda Jan 28 '19 at 08:10
Do you have a link to any tutorial/source that explains the model retraining in action? – Japesh Methuku Apr 27 '21 at 12:02
1

@JapeshMethuku No, since I am over time more in PyTorch, but you may try https://keras.io/guides /transfer_learning/ – prosti Apr 27 '21 at 18:10
@prosti thanks! but I didn't refer to transfer learning when I told retraining. – Japesh Methuku Apr 28 '21 at 00:35
Is there any big difference? – prosti Apr 28 '21 at 06:45

How to retrain/update keras model?

1 Answers1