This may be a basic question, but what is the recommended way to use Keras when I need to do some modifications to it?

[KyotoSunshine]

For example, suppose I want to use keras, but I want to try a different weight initialization method, which is not included in keras. How am I supposed to use Keras in this case (without modifying the source code of keras itself) ?

[mthrok]

You can define your custom function as Theano function and pass it to model.

See my answer in #369 for objective function case. I think it's same for initialization function.

[KyotoSunshine]

Is it really same? Here is the code that gets executed on the value of init parameter:

    self.init = initializations.get(init)

It seems that the value must be a string rather than a function, isn't it?

[KyotoSunshine]

I did it like the following, please let me know if there is a better method!

def custom_initialization(shape):
    fan_in, fan_out = keras.initializations.get_fans(shape)
    return keras.utils.theano_utils.sharedX(myfunction(shape, fan_in,_fan_out))

keras.initializations.custom_initialization = custom_initialization
...
model.add(Dense(784, 128, init='custom_initialization'))

[mthrok]

Did you see the inside of the initializations.get ? https://github.com/fchollet/keras/blob/master/keras/utils/generic_utils.py#L7
It checks if the given argument is string and, if it's string, it looks for a pre-defined function, and if not, look for global variable table and use the identifier you give.

So this works and I verified it.

import theano
import theano.tensor as T
from keras.utils.theano_utils import sharedX

def custom_initialization(shape, scale=0.05):
    '''just another normal'''
    return sharedX(np.random.randn(*shape) * scale)

model = Sequential()
model.add(Convolution2D(32, 1, 3, 3, border_mode='full', init=custom_initialization))

[KyotoSunshine]

Ah, I see. Looks very nice!