custom_layers.py 文件源码

def __init__(self, incoming, num_units, max_steps, peepholes=False, mask_input=None, **kwargs):
        """
        initialization
        :param incoming: bidirectional mLSTM for passane
        :param num_units:
        :param max_steps: max num steps to generate answer words, can be tensor scalar variable
        :param peepholes:
        :param mask_input: passage's length mask
        :param kwargs:
        """
        super(AnsPointerLayer, self).__init__(incoming, num_units, peepholes=peepholes,
                                              precompute_input=False, mask_input=mask_input,
                                              only_return_final=False, **kwargs)
        self.max_steps = max_steps
        # initializes attention weights
        input_shape = self.input_shapes[0]
        num_inputs = np.prod(input_shape[2:])
        self.V_pointer = self.add_param(init.Normal(0.1), (num_inputs, num_units), 'V_pointer')
        # doesn't need transpose
        self.v_pointer = self.add_param(init.Normal(0.1), (num_units, 1), 'v_pointer')
        self.W_a_pointer = self.add_param(init.Normal(0.1), (num_units, num_units), 'W_a_pointer')
        self.b_a_pointer = self.add_param(init.Constant(0.), (num_units, ), 'b_a_pointer')
        c_pointer = theano.shared(np.array([0.], dtype='float32'), name='c_pointer', broadcastable=(True, ))
        self.c_pointer = self.add_param(c_pointer, (1,), 'c_pointer')