networks.py 文件源码

def build_BiGRU(incoming, num_units, mask=None, grad_clipping=0, precompute_input=True, dropout=True, in_to_out=False):
    # construct the forward and backward grus. Now, Ws are initialized by Glorot initializer with default arguments.
    # Need to try other initializers for specific tasks.

    # dropout for incoming
    if dropout:
        incoming = lasagne.layers.DropoutLayer(incoming, p=0.5)

    # according to Jozefowicz et al.(2015), init bias of forget gate to 1.
    resetgate_forward = Gate(W_in=lasagne.init.GlorotUniform(), W_hid=lasagne.init.GlorotUniform(),
                             W_cell=lasagne.init.Uniform(range=0.1), b=lasagne.init.Constant(1.))
    updategate_forward = Gate(W_in=lasagne.init.GlorotUniform(), W_hid=lasagne.init.GlorotUniform(),
                              W_cell=lasagne.init.Uniform(range=0.1))
    # now use tanh for nonlinear function of hidden gate
    hidden_forward = Gate(W_in=lasagne.init.GlorotUniform(), W_hid=lasagne.init.GlorotUniform(), W_cell=None,
                          nonlinearity=nonlinearities.tanh)
    gru_forward = lasagne.layers.GRULayer(incoming, num_units, mask_input=mask, grad_clipping=grad_clipping,
                                          precompute_input=precompute_input,
                                          resetgate=resetgate_forward, updategate=updategate_forward,
                                          hidden_update=hidden_forward, name='forward')

    # according to Jozefowicz et al.(2015), init bias of forget gate to 1.
    resetgate_backward = Gate(W_in=lasagne.init.GlorotUniform(), W_hid=lasagne.init.GlorotUniform(),
                              W_cell=lasagne.init.Uniform(range=0.1), b=lasagne.init.Constant(1.))
    updategate_backward = Gate(W_in=lasagne.init.GlorotUniform(), W_hid=lasagne.init.GlorotUniform(),
                               W_cell=lasagne.init.Uniform(range=0.1))
    # now use tanh for nonlinear function of hidden gate
    hidden_backward = Gate(W_in=lasagne.init.GlorotUniform(), W_hid=lasagne.init.GlorotUniform(), W_cell=None,
                           nonlinearity=nonlinearities.tanh)
    gru_backward = lasagne.layers.GRULayer(incoming, num_units, mask_input=mask, grad_clipping=grad_clipping,
                                           precompute_input=precompute_input, backwards=True,
                                           resetgate=resetgate_backward, updategate=updategate_backward,
                                           hidden_update=hidden_backward, name='backward')

    # concatenate the outputs of forward and backward GRUs to combine them.
    concat = lasagne.layers.concat([gru_forward, gru_backward], axis=2, name="bi-gru")

    # dropout for output
    if dropout:
        concat = lasagne.layers.DropoutLayer(concat, p=0.5)

    if in_to_out:
        concat = lasagne.layers.concat([concat, incoming], axis=2)

    # the shape of BiRNN output (concat) is (batch_size, input_length, 2 * num_hidden_units)
    return concat