network.py 文件源码

def __init__(self, input_shape, output_dim, hidden_dim, hidden_nonlinearity=LN.rectify,
                 output_nonlinearity=None, name=None, input_var=None, input_layer=None):
        if input_layer is None:
            l_in = L.InputLayer(shape=(None, None) + input_shape, input_var=input_var, name="input")
        else:
            l_in = input_layer
        l_step_input = L.InputLayer(shape=(None,) + input_shape)
        l_step_prev_hidden = L.InputLayer(shape=(None, hidden_dim))
        l_gru = GRULayer(l_in, num_units=hidden_dim, hidden_nonlinearity=hidden_nonlinearity,
                         hidden_init_trainable=False)
        l_gru_flat = L.ReshapeLayer(
            l_gru, shape=(-1, hidden_dim)
        )
        l_output_flat = L.DenseLayer(
            l_gru_flat,
            num_units=output_dim,
            nonlinearity=output_nonlinearity,
        )
        l_output = OpLayer(
            l_output_flat,
            op=lambda flat_output, l_input:
            flat_output.reshape((l_input.shape[0], l_input.shape[1], -1)),
            shape_op=lambda flat_output_shape, l_input_shape:
            (l_input_shape[0], l_input_shape[1], flat_output_shape[-1]),
            extras=[l_in]
        )
        l_step_hidden = l_gru.get_step_layer(l_step_input, l_step_prev_hidden)
        l_step_output = L.DenseLayer(
            l_step_hidden,
            num_units=output_dim,
            nonlinearity=output_nonlinearity,
            W=l_output_flat.W,
            b=l_output_flat.b,
        )

        self._l_in = l_in
        self._hid_init_param = l_gru.h0
        self._l_gru = l_gru
        self._l_out = l_output
        self._l_step_input = l_step_input
        self._l_step_prev_hidden = l_step_prev_hidden
        self._l_step_hidden = l_step_hidden
        self._l_step_output = l_step_output