blocks.py 文件源码

def recurrent(network, mask_in, num_rec_units, num_layers, dropout,
              bidirectional, nonlinearity):

    if nonlinearity != 'LSTM':
        nl = getattr(lnn.nonlinearities, nonlinearity)

        def add_layer(prev_layer, **kwargs):
            return lnn.layers.RecurrentLayer(
                prev_layer, num_units=num_rec_units, mask_input=mask_in,
                nonlinearity=nl,
                W_in_to_hid=lnn.init.GlorotUniform(),
                W_hid_to_hid=lnn.init.Orthogonal(gain=np.sqrt(2) / 2),
                **kwargs)

    else:
        def add_layer(prev_layer, **kwargs):
            return lnn.layers.LSTMLayer(
                prev_layer, num_units=num_rec_units, mask_input=mask_in,
                **kwargs
            )

    fwd = network
    for i in range(num_layers):
        fwd = add_layer(fwd, name='rec_fwd_{}'.format(i))
        if dropout > 0.:
            fwd = lnn.layers.DropoutLayer(fwd, p=dropout)

    if not bidirectional:
        return network

    bck = network
    for i in range(num_layers):
        bck = add_layer(bck, name='rec_bck_{}'.format(i), backwards=True)
        if dropout > 0:
            bck = lnn.layers.DropoutLayer(bck, p=dropout)

    # combine the forward and backward recurrent layers...
    network = lnn.layers.ConcatLayer([fwd, bck], name='fwd + bck', axis=-1)
    return network