I was wondering if it was possible to save a partly trained Keras model and continue the training after loading the model again.

The reason for this is that I will have more training data in the future and I do not want to retrain the whole model again.

The functions which I am using are:

#Partly train model
model.fit(first_training, first_classes, batch_size=32, nb_epoch=20)

#Save partly trained model
model.save('partly_trained.h5')

#Load partly trained model
from keras.models import load_model
model = load_model('partly_trained.h5')

#Continue training
model.fit(second_training, second_classes, batch_size=32, nb_epoch=20)

Edit 1: added fully working example

With the first dataset after 10 epochs the loss of the last epoch will be 0.0748 and the accuracy 0.9863.

After saving, deleting and reloading the model the loss and accuracy of the model trained on the second dataset will be 0.1711 and 0.9504 respectively.

Is this caused by the new training data or by a completely re-trained model?

"""
Model by: http://machinelearningmastery.com/
"""
# load (downloaded if needed) the MNIST dataset
import numpy
from keras.datasets import mnist
from keras.models import Sequential
from keras.layers import Dense
from keras.utils import np_utils
from keras.models import load_model
numpy.random.seed(7)

def baseline_model():
    model = Sequential()
    model.add(Dense(num_pixels, input_dim=num_pixels, init='normal', activation='relu'))
    model.add(Dense(num_classes, init='normal', activation='softmax'))
    model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])
    return model

if __name__ == '__main__':
    # load data
    (X_train, y_train), (X_test, y_test) = mnist.load_data()

    # flatten 28*28 images to a 784 vector for each image
    num_pixels = X_train.shape[1] * X_train.shape[2]
    X_train = X_train.reshape(X_train.shape[0], num_pixels).astype('float32')
    X_test = X_test.reshape(X_test.shape[0], num_pixels).astype('float32')
    # normalize inputs from 0-255 to 0-1
    X_train = X_train / 255
    X_test = X_test / 255
    # one hot encode outputs
    y_train = np_utils.to_categorical(y_train)
    y_test = np_utils.to_categorical(y_test)
    num_classes = y_test.shape[1]

    # build the model
    model = baseline_model()

    #Partly train model
    dataset1_x = X_train[:3000]
    dataset1_y = y_train[:3000]
    model.fit(dataset1_x, dataset1_y, nb_epoch=10, batch_size=200, verbose=2)

    # Final evaluation of the model
    scores = model.evaluate(X_test, y_test, verbose=0)
    print("Baseline Error: %.2f%%" % (100-scores[1]*100))

    #Save partly trained model
    model.save('partly_trained.h5')
    del model

    #Reload model
    model = load_model('partly_trained.h5')

    #Continue training
    dataset2_x = X_train[3000:]
    dataset2_y = y_train[3000:]
    model.fit(dataset2_x, dataset2_y, nb_epoch=10, batch_size=200, verbose=2)
    scores = model.evaluate(X_test, y_test, verbose=0)
    print("Baseline Error: %.2f%%" % (100-scores[1]*100))

reduce_lr = ReduceLROnPlateau(monitor='loss', factor=lr_reduction_factor,
                              patience=patience, min_lr=min_lr, verbose=1)

for the pretrained model, whereby the original learning rate starts at 0.0003 and during pre-training it is reduced to the min_learning rate, which is 0.000003

I just copied that line over to the script which uses the pre-trained model and got really bad accuracies. Until I noticed that the last learning rate of the pretrained model was the min learning rate, i.e. 0.000003. And if I start with that learning rate, I get exactly the same accuracies to start with as the output of the pretrained model - which makes sense, as starting with a learning rate that is 100 times bigger than the last learning rate used in the pretrained model will result in a huge overshoot of GD and hence in heavily decreased accuracies.

import tensorflow as tf
from tensorflow import keras
mnist = tf.keras.datasets.mnist

(x_train, y_train),(x_test, y_test) = mnist.load_data()
x_train, x_test = x_train / 255.0, x_test / 255.0

def create_model():
  model = tf.keras.models.Sequential([
    tf.keras.layers.Flatten(input_shape=(28, 28)),
    tf.keras.layers.Dense(512, activation=tf.nn.relu),  
    tf.keras.layers.Dropout(0.2),
    tf.keras.layers.Dense(10, activation=tf.nn.softmax)
    ])

  model.compile(optimizer='adam', loss='sparse_categorical_crossentropy',metrics=['accuracy'])
  return model

# Create a basic model instance
model=create_model()
model.fit(x_train, y_train, epochs = 10, validation_data = (x_test,y_test),verbose=1)​

# saving the model in tensorflow format
model.save('./MyModel_tf',save_format='tf')


# loading the saved model
loaded_model = tf.keras.models.load_model('./MyModel_tf')

# retraining the model
loaded_model.fit(x_train, y_train, epochs = 10, validation_data = (x_test,y_test),verbose=1)​

All above helps, you must resume from same learning rate() as the LR when the model and weights were saved. Set it directly on the optimizer.

Note that improvement from there is not guaranteed, because the model may have reached the local minimum, which may be global. There is no point to resume a model in order to search for another local minimum, unless you intent to increase the learning rate in a controlled fashion and nudge the model into a possibly better minimum not far away.