def _initialize_theta(self):
filter_shape = self.filter_shape
image_shape = self.image_shape
poolsize = self.poolsize
conv_in = np.prod(filter_shape[1:])
conv_out = filter_shape[0] * np.prod(filter_shape[2:])
pool_out = conv_out / poolsize**2
conv_map_size = image_shape[-1] - filter_shape[-1] + 1
assert conv_map_size > 0
pool_map_size = int(conv_map_size / poolsize)
assert pool_map_size > 0
self.conv_w = shared(
nn_random_paramters(conv_in, conv_out, shape=filter_shape))
self.conv_b = shared(
nn_random_paramters(conv_in, conv_out,
shape=(filter_shape[0], 1, 1)))
self.pool_w = shared(
nn_random_paramters(conv_out, pool_out,
shape=(filter_shape[0], 1, 1)))
self.pool_b = shared(
nn_random_paramters(conv_out, pool_out,
shape=(filter_shape[0], 1, 1)))
self.output_shape = (image_shape[0], filter_shape[0],
pool_map_size, pool_map_size)
return [self.conv_w, self.conv_b, self.pool_w, self.pool_b]
def _initialize_theta(self):
"""Return a set value which can be used to initialize weights.
Override this function to use other initial value.
"""
return nn_random_paramters(self.n_in+1, self.n_out)
