network.py 文件源码

def __init__(self, incoming, num_units, hidden_nonlinearity,
                 gate_nonlinearity=LN.sigmoid, name=None,
                 W_init=LI.GlorotUniform(), b_init=LI.Constant(0.),
                 hidden_init=LI.Constant(0.), hidden_init_trainable=True):

        if hidden_nonlinearity is None:
            hidden_nonlinearity = LN.identity

        if gate_nonlinearity is None:
            gate_nonlinearity = LN.identity

        super(GRULayer, self).__init__(incoming, name=name)

        input_shape = self.input_shape[2:]

        input_dim = ext.flatten_shape_dim(input_shape)
        # self._name = name
        # Weights for the initial hidden state
        self.h0 = self.add_param(hidden_init, (num_units,), name="h0", trainable=hidden_init_trainable,
                                 regularizable=False)
        # Weights for the reset gate
        self.W_xr = self.add_param(W_init, (input_dim, num_units), name="W_xr")
        self.W_hr = self.add_param(W_init, (num_units, num_units), name="W_hr")
        self.b_r = self.add_param(b_init, (num_units,), name="b_r", regularizable=False)
        # Weights for the update gate
        self.W_xu = self.add_param(W_init, (input_dim, num_units), name="W_xu")
        self.W_hu = self.add_param(W_init, (num_units, num_units), name="W_hu")
        self.b_u = self.add_param(b_init, (num_units,), name="b_u", regularizable=False)
        # Weights for the cell gate
        self.W_xc = self.add_param(W_init, (input_dim, num_units), name="W_xc")
        self.W_hc = self.add_param(W_init, (num_units, num_units), name="W_hc")
        self.b_c = self.add_param(b_init, (num_units,), name="b_c", regularizable=False)
        self.gate_nonlinearity = gate_nonlinearity
        self.num_units = num_units
        self.nonlinearity = hidden_nonlinearity