capsule.py

import numpy as np
import theano
import theano.tensor as T

class Capsule():

    def __init__(self, name, n_hidden, n_output, num_caps):

        self.name = name
        bias = np.asarray([0,0, 3 * num_caps,1,1,1], dtype=theano.config.floatX)/ num_caps
        self.params = [self.init_capsule_weight(n_hidden),
            theano.shared(bias, borrow=True)]


    def init_capsule_weight(self, n_hidden_l3):    

        l3_to_center = 0.05*np.asarray(
            np.random.uniform(
                low=-4 * np.sqrt(6. / 6+n_hidden_l3),
                high=4 * np.sqrt(6. / 6+n_hidden_l3),
                size=(n_hidden_l3, 3)
            ), dtype=theano.config.floatX)
        l3_to_radius = 0.0005*np.asarray( np.random.uniform(
                low=-4 * np.sqrt(6. / 6+n_hidden_l3),
                high=4 * np.sqrt(6. / 6+n_hidden_l3),
                size=(n_hidden_l3, 3)
            ), dtype=theano.config.floatX)
        return theano.shared(np.concatenate((l3_to_center, l3_to_radius), 1))


        #return theano.shared(0.07*np.asarray(
        #        np.random.uniform(
        #            low=-4 * np.sqrt(6. / n_output+n_hidden_l3),
        #            high=4 * np.sqrt(6. / n_output+n_hidden_l3),
        #            size=(n_hidden_l3, n_output)
        #        ), dtype=theano.config.floatX),borrow=True)