link
Keras Chinese document

One 、 How to save Keras Model ？

1. preservation / Load the whole model （ structure + The weight + Optimizer status ）

Not recommended pickle or cPickle To preserve Keras Model .

You can use model.save(filepath) take Keras Save the model to a single file HDF5 In file , The file will contain ：

• Structure of model , Allows you to recreate the model
• Model weight Training configuration item （ Loss function , Optimizer ）
• Optimizer status , Allow you to continue training exactly where you ended last time .

You can use keras.models.load_model(filepath) Re instantiate the model .load_model You will also be responsible for compiling the model using the saved training configuration items （ Unless the model has never been compiled ）.

Example ：

from keras.models import load_model

model.save('my_model.h5')  #  establish  HDF5  file  'my_model.h5'
del model  #  Delete existing model 

#  Return a compiled model 
#  Same as the one before 
model = load_model('my_model.h5')

See also How to install HDF5 or h5py In the Keras Save my model in ？, See how to install h5py Explanation .

2. Save only / Load the structure of the model

If you only need to save the structure of the model , Instead of its weight or training configuration item , You can do the following ：

#  Save as  JSON
json_string = model.to_json()

#  Save as  YAML
yaml_string = model.to_yaml()

Generated JSON/YAML Files are human readable , You can also edit it manually if necessary .
You can build a new model from this data ：

#  from  JSON  Reconstruction model ：
from keras.models import model_from_json
model = model_from_json(json_string)

#  from  YAML  Reconstruction model ：
from keras.models import model_from_yaml
model = model_from_yaml(yaml_string)

3. Save only / Load the weight of the model

If you only need Model weight , You can use the following code to HDF5 Format to save .
Please note that , We need to install... First HDF5 and Python library h5py, They are not included in Keras in .

model.save_weights('my_model_weights.h5')

Suppose you have code for instantiating the model , Then the saved weights can be loaded into the model with the same structure ：

model.load_weights('my_model_weights.h5')

If you need to load weights into different structures （ There are some common layers ） In the model of , For example, fine tune or transfer learning , You can load the weight by the name of the layer ：

model.load_weights('my_model_weights.h5', by_name=True)

Example ：

"""
 Suppose the original model is as follows ：
    model = Sequential()
    model.add(Dense(2, input_dim=3, name='dense_1'))
    model.add(Dense(3, name='dense_2'))
    ...
    model.save_weights(fname)
"""

#  The new model 
model = Sequential()
model.add(Dense(2, input_dim=3, name='dense_1'))  #  Will be loaded 
model.add(Dense(10, name='new_dense'))  #  Will not be loaded 

#  Load weights from the first model ; It will only affect the first layer ,dense_1
model.load_weights(fname, by_name=True)

4. Working with custom layers in a saved model （ Or other custom objects ）

If the model to be loaded contains custom layers or other custom classes or functions , You can use the custom_objects Parameters are passed to the loading mechanism ：

from keras.models import load_model
#  Suppose your model contains a  AttentionLayer  Class 
model = load_model('my_model.h5', custom_objects={
    'AttentionLayer': AttentionLayer})

perhaps , You can use Custom object scope ：

from keras.utils import CustomObjectScope

with CustomObjectScope({
    'AttentionLayer': AttentionLayer}):
    model = load_model('my_model.h5')

Processing and of custom objects load_model, model_from_json, model_from_yaml They work the same way ：

from keras.models import model_from_json
model = model_from_json(json_string, custom_objects={
    'AttentionLayer': AttentionLayer})