RNNCell.py 文件源码

def __call__(self, x, state_prev, scope=None):
        with tf.variable_scope(scope or type(self).__name__):

            h_prev, c_prev = state_prev

            # Check if the input size exist.
            input_size = x.shape.with_rank(2)[1].value
            if input_size is None:
                raise ValueError("Expecting input_size to be set.")

            ### get weights for concated tensor.
            W_shape = (input_size, self.output_size)
            U_shape = (self.output_size, self.output_size)
            b_shape = (self.output_size,)
            Wi = tf.get_variable(name='Wi', shape=W_shape)
            Wj = tf.get_variable(name='Wj', shape=W_shape)
            Wf = tf.get_variable(name='Wf', shape=W_shape)
            Wo = tf.get_variable(name='Wo', shape=W_shape)
            Ui = tf.get_variable(name='Ui', shape=U_shape,
                                 initializer=TFCommon.Initializer.random_orthogonal_initializer())
            Uj = tf.get_variable(name='Uj', shape=U_shape,
                                 initializer=TFCommon.Initializer.random_orthogonal_initializer())
            Uf = tf.get_variable(name='Uf', shape=U_shape,
                                 initializer=TFCommon.Initializer.random_orthogonal_initializer())
            Uo = tf.get_variable(name='Uo', shape=U_shape,
                                 initializer=TFCommon.Initializer.random_orthogonal_initializer())
            bi = tf.get_variable(name='bi', shape=b_shape,
                                 initializer=tf.constant_initializer(0.0))
            bj = tf.get_variable(name='bj', shape=b_shape,
                                 initializer=tf.constant_initializer(0.0))
            bf = tf.get_variable(name='bf', shape=b_shape,
                                 initializer=tf.constant_initializer(self.__forget_bias)) # forget gate bias := 1
            bo = tf.get_variable(name='bo', shape=b_shape,
                                 initializer=tf.constant_initializer(0.0))

            ### calculate gates and input's info
            i = tf.tanh(tf.matmul(x, Wi) + tf.matmul(h_prev, Ui) + bi)
            j = self.__gate_activation(tf.matmul(x, Wj) + tf.matmul(h_prev, Uj) + bj)
            f = self.__gate_activation(tf.matmul(x, Wf) + tf.matmul(h_prev, Uf) + bf)
            o = tf.tanh(tf.matmul(x, Wo) + tf.matmul(h_prev, Uo) + bo)

            ### calculate candidate
            new_c = f * c_prev + i * j

            ### final cal
            new_h = o * tf.tanh(new_c)

            return new_h, LSTMStateTuple(new_h, new_c)