How to save Scikit-Learn-Keras Model into a Persistence File (pickle/hd5/json/yaml)

Question

I have the following code, using Keras Scikit-Learn Wrapper:

from keras.models import Sequential
from sklearn import datasets
from keras.layers import Dense
from sklearn.model_selection import train_test_split
from keras.wrappers.scikit_learn import KerasClassifier
from sklearn.model_selection import StratifiedKFold
from sklearn.model_selection import cross_val_score
from sklearn import preprocessing
import pickle
import numpy as np
import json

def classifier(X, y):
    """
    Description of classifier
    """
    NOF_ROW, NOF_COL =  X.shape

    def create_model():
        # create model
        model = Sequential()
        model.add(Dense(12, input_dim=NOF_COL, init='uniform', activation='relu'))
        model.add(Dense(6, init='uniform', activation='relu'))
        model.add(Dense(1, init='uniform', activation='sigmoid'))
        # Compile model
        model.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])
        return model

    # evaluate using 10-fold cross validation
    seed = 7
    np.random.seed(seed)
    model = KerasClassifier(build_fn=create_model, nb_epoch=150, batch_size=10, verbose=0)
    return model


def main():
    """
    Description of main
    """

    iris = datasets.load_iris()
    X, y = iris.data, iris.target
    X = preprocessing.scale(X)

    X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.4, random_state=0)
    model_tt = classifier(X_train, y_train)
    model_tt.fit(X_train,y_train)

    #--------------------------------------------------
    # This fail
    #-------------------------------------------------- 
    filename = 'finalized_model.sav'
    pickle.dump(model_tt, open(filename, 'wb'))
    # load the model from disk
    loaded_model = pickle.load(open(filename, 'rb'))
    result = loaded_model.score(X_test, Y_test)
    print(result)

    #--------------------------------------------------
    # This also fail
    #--------------------------------------------------
    # from keras.models import load_model       
    # model_tt.save('test_model.h5')


    #--------------------------------------------------
    # This works OK 
    #-------------------------------------------------- 
    # print model_tt.score(X_test, y_test)
    # print model_tt.predict_proba(X_test)
    # print model_tt.predict(X_test)


# Output of predict_proba
# 2nd column is the probability that the prediction is 1
# this value is used as final score, which can be used
# with other method as comparison
# [   [ 0.25311464  0.74688536]
#     [ 0.84401423  0.15598579]
#     [ 0.96047372  0.03952631]
#     ...,
#     [ 0.25518912  0.74481088]
#     [ 0.91467732  0.08532269]
#     [ 0.25473493  0.74526507]]

# Output of predict
# [[1]
# [0]
# [0]
# ...,
# [1]
# [0]
# [1]]


if __name__ == '__main__':
    main()

As stated in the code there it fails at this line:

pickle.dump(model_tt, open(filename, 'wb'))

With this error:

pickle.PicklingError: Can't pickle <function create_model at 0x101c09320>: it's not found as __main__.create_model

How can I get around it?

Answer 1

Edit 1 : Original answer about saving model

With HDF5 :

# saving model
json_model = model_tt.model.to_json()
open('model_architecture.json', 'w').write(json_model)
# saving weights
model_tt.model.save_weights('model_weights.h5', overwrite=True)


# loading model
from keras.models import model_from_json

model = model_from_json(open('model_architecture.json').read())
model.load_weights('model_weights.h5')

# dont forget to compile your model
model.compile(loss='binary_crossentropy', optimizer='adam')

Edit 2 : full code example with iris dataset

# Train model and make predictions
import numpy
import pandas
from keras.models import Sequential, model_from_json
from keras.layers import Dense
from keras.utils import np_utils
from sklearn import datasets
from sklearn import preprocessing
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import LabelEncoder

# fix random seed for reproducibility
seed = 7
numpy.random.seed(seed)

# load dataset
iris = datasets.load_iris()
X, Y, labels = iris.data, iris.target, iris.target_names
X = preprocessing.scale(X)

# encode class values as integers
encoder = LabelEncoder()
encoder.fit(Y)
encoded_Y = encoder.transform(Y)

# convert integers to dummy variables (i.e. one hot encoded)
y = np_utils.to_categorical(encoded_Y)

def build_model():
    # create model
    model = Sequential()
    model.add(Dense(4, input_dim=4, init='normal', activation='relu'))
    model.add(Dense(3, init='normal', activation='sigmoid'))
    model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])
    return model

def save_model(model):
    # saving model
    json_model = model.to_json()
    open('model_architecture.json', 'w').write(json_model)
    # saving weights
    model.save_weights('model_weights.h5', overwrite=True)

def load_model():
    # loading model
    model = model_from_json(open('model_architecture.json').read())
    model.load_weights('model_weights.h5')
    model.compile(loss='categorical_crossentropy', optimizer='adam')
    return model


X_train, X_test, Y_train, Y_test = train_test_split(X, y, test_size=0.3, random_state=seed)

# build
model = build_model()
model.fit(X_train, Y_train, nb_epoch=200, batch_size=5, verbose=0)

# save
save_model(model)

# load
model = load_model()

# predictions
predictions = model.predict_classes(X_test, verbose=0)
print(predictions)
# reverse encoding
for pred in predictions:
    print(labels[pred])

Please note that I used Keras only, not the wrapper. It only add some complexity in something simple. Also code is voluntarily not refactored so you can have the whole picture.

Also, you said you want to output 1 or 0. It is not possible in this dataset because you have 3 output dims and classes (Iris-setosa, Iris-versicolor, Iris-virginica). If you had only 2 classes then your output dim and classes would be 0 or 1 using sigmoid output fonction.

Answer 2

Just adding to gaarv's answer - If you don't require the separation between the model structure (model.to_json()) and the weights (model.save_weights()), you can use one of the following:

• Use the built-in keras.models.save_model and 'keras.models.load_model` that store everything together in a hdf5 file.
• Use pickle to serialize the Model object (or any class that contains references to it) into file/network/whatever..
  Unfortunetaly, Keras doesn't support pickle by default. You can use my patchy solution that adds this missing feature. Working code is here: http://zachmoshe.com/2017/04/03/pickling-keras-models.html

Answer 3

Another great alternative is to use callbacks when you fit your model. Specifically the ModelCheckpoint callback, like this:

from keras.callbacks import ModelCheckpoint
#Create instance of ModelCheckpoint
chk = ModelCheckpoint("myModel.h5", monitor='val_loss', save_best_only=False)
#add that callback to the list of callbacks to pass
callbacks_list = [chk]
#create your model
model_tt = KerasClassifier(build_fn=create_model, nb_epoch=150, batch_size=10)
#fit your model with your data. Pass the callback(s) here
model_tt.fit(X_train,y_train, callbacks=callbacks_list)

This will save your training each epoch to the myModel.h5 file. This provides great benefits, as you are able to stop your training when you desire (like when you see it has started to overfit), and still retain the previous training.

Note that this saves both the structure and weights in the same hdf5 file (as showed by Zach), so you can then load you model using keras.models.load_model.

If you want to save only your weights separately, you can then use the save_weights_only=True argument when instantiating your ModelCheckpoint, enabling you to load your model as explained by Gaarv. Extracting from the docs:

save_weights_only: if True, then only the model's weights will be saved (model.save_weights(filepath)), else the full model is saved (model.save(filepath)).

Answer 4

The accepted answer is too complicated. You can fully save and restore every aspect of your model in a .h5 file. Straight from the Keras FAQ:

You can use model.save(filepath) to save a Keras model into a single HDF5 file which will contain:

• the architecture of the model, allowing to re-create the model
• the weights of the model
• the training configuration (loss, optimizer)
• the state of the optimizer, allowing to resume training exactly where you left off.

You can then use keras.models.load_model(filepath) to reinstantiate your model. load_model will also take care of compiling the model using the saved training configuration (unless the model was never compiled in the first place).

And the corresponding code:

from keras.models import load_model

model.save('my_model.h5')  # creates a HDF5 file 'my_model.h5'
del model  # deletes the existing model

# returns a compiled model identical to the previous one
model = load_model('my_model.h5')

Answer 5

In case your keras wrapper model is in a scikit pipeline, you save steps in the pipeline separately.

import joblib
from sklearn.pipeline import Pipeline
from tensorflow import keras

# pass the create_cnn_model function into wrapper
cnn_model = keras.wrappers.scikit_learn.KerasClassifier(build_fn=create_cnn_model)

# create pipeline
cnn_model_pipeline_estimator = Pipeline([
    ('preprocessing_pipeline', pipeline_estimator),
    ('clf', cnn_model)
])

# train model
final_model = cnn_model_pipeline_estimator.fit(
X, y, clf__batch_size=32, clf__epochs=15)

# collect the preprocessing pipeline & model seperately
pipeline_estimator = final_model.named_steps['preprocessing_pipeline']
clf = final_model.named_steps['clf']

# store pipeline and model seperately
joblib.dump(pipeline_estimator, open('path/to/pipeline.pkl', 'wb'))
clf.model.save('path/to/model.h5')

# load pipeline and model
pipeline_estimator = joblib.load('path/to/pipeline.pxl')
model = keras.models.load_model('path/to/model.h5')

new_example = [[...]]

# transform new data with pipeline & use model for prediction
transformed_data = pipeline_estimator.transform(new_example)
prediction = model.predict(transformed_data)