如何将 Scikit-Learn-Keras 模型保存到持久性文件 (pickle/hd5/json/yaml)

Question

我有以下代码，使用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()

如代码中所述，它在这一行失败：

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

出现此错误：

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

我怎样才能绕过它？

Answer 1

接受的答案太复杂了。您可以在.h5 文件中完全保存和恢复模型的各个方面。直接来自Keras FAQ：

您可以使用model.save(filepath) 将 Keras 模型保存到单个 HDF5 文件将包含：

• 模型的架构，允许重新创建模型
• 模型的权重
• 训练配置（损失、优化器）
• 优化器的状态，允许在您停止的地方恢复训练。

然后您可以使用keras.models.load_model(filepath) 重新实例化您的模型。 load_model 还将负责使用保存的训练配置编译模型（除非模型一开始就从未编译过）。

以及对应的代码：

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 2

如果您的 keras 包装器模型位于 scikit 管道中，您可以单独保存管道中的步骤。

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)

Answer 3

另一个很好的选择是在你fit你的模型时使用callbacks。特别是ModelCheckpoint 回调，像这样：

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)

这会将您的训练每个时期保存到myModel.h5 文件中。这提供了很大的好处，因为您可以在需要时停止训练（例如当您看到它开始过度拟合时），并且仍然保留之前的训练。

请注意，这会将结构和权重保存在同一个 hdf5 文件中（如 Zach 所示），因此您可以使用 keras.models.load_model 加载模型。

如果您只想单独保存权重，则可以在实例化 ModelCheckpoint 时使用 save_weights_only=True 参数，使您能够按照 Gaarv 的说明加载模型。提取自docs：

save_weights_only：如果为 True，则仅保存模型的权重 (model.save_weights(filepath))，否则保存完整模型 (model.save(filepath))。

Answer 4

只是添加到 gaarv 的答案 - 如果您不需要分离模型结构 (model.to_json()) 和权重 (model.save_weights())，您可以使用以下方法之一：

• 使用内置的 keras.models.save_model 和 'keras.models.load_model` 将所有内容一起存储在 hdf5 文件中。
• 使用 pickle 将 Model 对象（或任何包含对它的引用的类）序列化为文件/网络/任何东西..
  不幸的是，Keras 默认不支持 pickle。您可以使用 我的补丁解决方案添加了这个缺失的功能。工作代码是 在这里：http://zachmoshe.com/2017/04/03/pickling-keras-models.html

Answer 5

编辑1：关于保存模型的原始答案

使用 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')

编辑 2 ：带有 iris 数据集的完整代码示例

# 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])

请注意，我只使用了 Keras，而不是包装器。它只会在简单的事情中增加一些复杂性。此外，代码是自愿的，不考虑因素，因此您可以了解整个情况。

另外，您说过要输出 1 或 0。在此数据集中这是不可能的，因为您有 3 个输出尺寸和类别（Iris-setosa、Iris-versicolor、Iris-virginica）。如果您只有 2 个类，那么使用 sigmoid 输出函数，您的输出 dim 和类将为 0 或 1。