sent_classify.py 文件源码

def build_RNN(architec, layer_input, layer_mask, num_units, grad_clipping):
    def build_GRU(reset_input):
        resetgate = Gate(W_in=lasagne.init.GlorotUniform(), W_hid=lasagne.init.GlorotUniform(), W_cell=None)

        updategate = Gate(W_in=lasagne.init.GlorotUniform(), W_hid=lasagne.init.GlorotUniform(), W_cell=None)

        hiden_update = Gate(W_in=lasagne.init.GlorotUniform(), W_hid=lasagne.init.GlorotUniform(), W_cell=None,
                            b=lasagne.init.Constant(0.), nonlinearity=nonlinearities.tanh)

        return GRULayer(layer_input, num_units, mask_input=layer_mask, grad_clipping=grad_clipping,
                        resetgate=resetgate, updategate=updategate, hidden_update=hiden_update,
                        reset_input=reset_input, only_return_final=True, p=0.5, name='GRU')

    def build_LSTM():
        ingate = Gate(W_in=lasagne.init.GlorotUniform(), W_hid=lasagne.init.GlorotUniform(),
                      W_cell=lasagne.init.Uniform(range=0.1))

        outgate = Gate(W_in=lasagne.init.GlorotUniform(), W_hid=lasagne.init.GlorotUniform(),
                       W_cell=lasagne.init.Uniform(range=0.1))
        # according to Jozefowicz et al.(2015), init bias of forget gate to 1.
        forgetgate = Gate(W_in=lasagne.init.GlorotUniform(), W_hid=lasagne.init.GlorotUniform(),
                          W_cell=lasagne.init.Uniform(range=0.1), b=lasagne.init.Constant(1.))
        # now use tanh for nonlinear function of cell, need to try pure linear cell
        cell = Gate(W_in=lasagne.init.GlorotUniform(), W_hid=lasagne.init.GlorotUniform(), W_cell=None,
                    b=lasagne.init.Constant(0.), nonlinearity=nonlinearities.tanh)

        return LSTMLayer(layer_input, num_units, mask_input=layer_mask, grad_clipping=grad_clipping,
                         ingate=ingate, forgetgate=forgetgate, cell=cell, outgate=outgate,
                         peepholes=False, nonlinearity=nonlinearities.tanh,
                         only_return_final=True, p=0.5, name='LSTM')

    def build_SGRU():
        resetgate_hidden = Gate(W_in=lasagne.init.GlorotUniform(), W_hid=lasagne.init.GlorotUniform(),
                         W_cell=lasagne.init.GlorotUniform())

        resetgate_input = Gate(W_in=lasagne.init.GlorotUniform(), W_hid=lasagne.init.GlorotUniform(),
                                W_cell=lasagne.init.GlorotUniform())

        updategate = Gate(W_in=lasagne.init.GlorotUniform(), W_hid=lasagne.init.GlorotUniform(),
                          W_cell=lasagne.init.GlorotUniform())

        hidden_update = Gate(W_in=lasagne.init.GlorotUniform(), W_hid=lasagne.init.GlorotUniform(), W_cell=None,
                            b=lasagne.init.Constant(0.), nonlinearity=nonlinearities.tanh)

        return SGRULayer(layer_input, num_units, mask_input=layer_mask, grad_clipping=grad_clipping,
                         resetgate_input=resetgate_input, resetgate_hidden=resetgate_hidden,
                         updategate=updategate, hidden_update=hidden_update,
                         only_return_final=True, p=0.5, name='SGRU')

    if architec == 'gru0':
        return build_GRU(False)
    elif architec == 'gru1':
        return build_GRU(True)
    elif architec == 'lstm':
        return build_LSTM()
    elif architec == 'sgru':
        return build_SGRU()
    else:
        raise ValueError('unkown architecture: %s' % architec)