h3 = lasagne.layers.RecurrentLayer(h2, num_classes, grad_clipping=grad_clip,
        nonlinearity=lasagne.nonlinearities.linear)
l_out = lasagne.layers.ReshapeLayer(h3, ((max_len, mbsz, num_classes)))

network_output = lasagne.layers.get_output(l_out)

h3 = lasagne.layers.RecurrentLayer(h2, num_classes, grad_clipping=grad_clip,
        nonlinearity=lasagne.nonlinearities.linear)
l_out = lasagne.layers.DimshuffleLayer(h3, (1, 0, 2))

network_output = lasagne.layers.get_output(l_out)

    # (batch size, max sequence length, number of features)
    l_in = lasagne.layers.InputLayer(shape=(None, None, inp_dim), input_var=input_var)
    # Mask as matrices of dimensionality (N_BATCH, MAX_LENGTH)
    l_mask = lasagne.layers.InputLayer(shape=(None, None), input_var=mask_var)
    # Allows arbitrary sizes
    batch_size, seq_len, _ = input_var.shape

    # RNN layers
    l_forward = lasagne.layers.GRULayer(
        l_in, num_hidden, mask_input=l_mask, grad_clipping=GRAD_CLIP)
    l_backward = lasagne.layers.GRULayer(
        l_in, num_hidden, mask_input=l_mask, grad_clipping=GRAD_CLIP, backwards=True)

    # Now, we'll concatenate the outputs to combine them.
    l_concat = lasagne.layers.ConcatLayer([l_forward, l_backward], axis=2)

    # Dense layers
    l_nfd1 = non_flattening_dense(l_concat, batch_size, seq_len, num_units=num_hidden,
        nonlinearity=lasagne.nonlinearities.rectify)
    l_nfd2 = non_flattening_dense(l_nfd1, batch_size, seq_len, num_units=num_hidden,
        nonlinearity=lasagne.nonlinearities.rectify)
    l_out = non_flattening_dense(l_nfd2, batch_size, seq_len, num_units=out_size+1, nonlinearity=lasagne.nonlinearities.linear)

    l_reshape =  lasagne.layers.DimshuffleLayer(l_out, (1, 0, 2))

    return l_reshape

def non_flattening_dense(l_in, batch_size, seq_len, *args, **kwargs):
    l_flat = lasagne.layers.ReshapeLayer(l_in, (-1, [2]))
    l_dense = lasagne.layers.DenseLayer(l_flat, *args, **kwargs)
    # temp_size = l_in.output_shape[1]
    # if not temp_size:
    #     temp_size = -1
    l_reshaped = lasagne.layers.ReshapeLayer(l_dense, (batch_size, seq_len, l_dense.output_shape[1]))
    return l_reshaped