avnet.py 文件源码

def create_model(substreams, mask_shape, mask_var, lstm_size=250, output_classes=26,
                 fusiontype='concat', w_init_fn=las.init.Orthogonal(), use_peepholes=True):

    gate_parameters = Gate(
        W_in=w_init_fn, W_hid=w_init_fn,
        b=las.init.Constant(0.))
    cell_parameters = Gate(
        W_in=w_init_fn, W_hid=w_init_fn,
        # Setting W_cell to None denotes that no cell connection will be used.
        W_cell=None, b=las.init.Constant(0.),
        # By convention, the cell nonlinearity is tanh in an LSTM.
        nonlinearity=tanh)

    l_mask = InputLayer(mask_shape, mask_var, 'mask')
    symbolic_seqlen_raw = l_mask.input_var.shape[1]

    # We'll combine the forward and backward layer output by summing.
    # Merge layers take in lists of layers to merge as input.
    if fusiontype == 'adasum':
        l_fuse = AdaptiveElemwiseSumLayer(substreams, name='adasum1')
    elif fusiontype == 'sum':
        l_fuse = ElemwiseSumLayer(substreams, name='sum1')
    elif fusiontype == 'concat':
        l_fuse = ConcatLayer(substreams, axis=-1, name='concat')

    f_lstm_agg, b_lstm_agg = create_blstm(l_fuse, l_mask, lstm_size, cell_parameters, gate_parameters, 'lstm_agg')
    l_sum2 = ElemwiseSumLayer([f_lstm_agg, b_lstm_agg], name='sum2')

    # reshape to (num_examples * seq_len, lstm_size)
    l_reshape3 = ReshapeLayer(l_sum2, (-1, lstm_size), name='reshape3')

    # Now, we can apply feed-forward layers as usual.
    # We want the network to predict a classification for the sequence,
    # so we'll use a the number of classes.
    l_softmax = DenseLayer(
        l_reshape3, num_units=output_classes,
        nonlinearity=las.nonlinearities.softmax, name='softmax')

    l_out = ReshapeLayer(l_softmax, (-1, symbolic_seqlen_raw, output_classes), name='output')

    return l_out, l_fuse