This repo is essentially a stripped down version of deepdish with some simple additions to make keeping track of keras models a bit easier. The entire code in generic/ is from deepdish, I just took out the relevant stuff.
The key addition here is that when you serialize a Keras model, the config and the weights are saved together. Though this is bad for things like net surgery, it's really useful when you have lots of configs and weights lying around and you can't make them fit together. Great for bottling up stuff for API launching, etc.
To save/load keras models, you can use:
from h5ify.keras_ext import save_model, load_model
# assume you have a Keras model
net = MyFancyModel()
save_model(net, './mynet.h5')
# loads uncompiled
nn = load_model('./mynet.h5')
# can load a compiled instance!
nn_other = load_model('./mynet.h5', 'sgd', 'mse', class_mode='categorical')To install it, just clone and do the usual python setup.py build and python setup.py install, with sudo's as needed. Or, just do pip install git+https://github.com/lukedeo/h5ify.git.