keras_pickle_wrapper

This small library exposes a KerasPickleWrapper class that allows keras models to be pickled, and even used across multiprocessing (or with a library like job_stream).

It won't be necessary if keras-team/keras#789 is ever properly resolved.

Installation

pip install keras-pickle-wrapper

Usage

import keras
import pickle
import numpy as np

from keras_pickle_wrapper import KerasPickleWrapper

ins = keras.layers.Input((2,))
x = ins
x = keras.layers.Dense(3)(x)
x = keras.layers.Dense(1)(x)
m = keras.models.Model(inputs=ins, outputs=x)
m.compile(loss='mse', optimizer='sgd')

# Wrap a compiled model
mw = KerasPickleWrapper(m)

# Calling the object returns the wrapped Keras model
mw().fit([[0,0], [0,1], [1,0], [1,1]], [[0], [1], [1], [0]])

# Pickle / unpickle the wrapper
data = pickle.dumps(mw)
mw2 = pickle.loads(data)
output_1 = mw().predict(np.asarray([[0, 0]]))
output_2 = mw2().predict(np.asarray([[0, 0]]))

# You can unload the object from memory as well
mw.unload()

# The object will remain unloaded until requested again
output_3 = mw().predict(np.asarray([[0, 0]]))

print("All outputs:")
print(output_1)
print(output_2)
print(output_3)

# If using tensorflow and pickling / unpickling a lot, be sure to clear the 
# session:
keras.backend.clear_session()

# When using a custom layer class, be sure to register it so the pickler works
class MyLayer(keras.layers.Layer):
    pass  # ...
KerasPickleWrapper.register(MyLayer)

If your model takes 1GB of RAM, the default approach should require 2GB additional RAM to encode, as it dumps to shared memory by default. To disable this, set KerasPickleWrapper.NO_SHM = True. Temporary files will then be written to the standard temporary directory. Using KerasPickleWrapper.unload(clear_session=True) prior to pickling combined with NO_SHM should eliminate excess memory consumption, but clears the session (session clearing only applies to Tensorflow at the moment).

Changelog

• 2018-7-17 - Fix for Python 2 in setup.py (FileNotFoundError not existing in Python 2.7), up to V1.0.5.
• 2017-9-19 - Fix for Python 2, up to V1.0.3.
• 2017-8-17 - Renamed to keras-pickle-wrapper because PyPI. V1.0.2.
• 2017-8-17 - Packaged up KerasWrapper for PyPI distribution.