util.py

def to_lodtensor(data, place):
    """convert tensor to lodtensor
    """
    seq_lens = [len(seq) for seq in data]
    cur_len = 0
    lod = [cur_len]
    for l in seq_lens:
        cur_len += l
        lod.append(cur_len)
    flattened_data = numpy.concatenate(data, axis=0).astype("int64")
    flattened_data = flattened_data.reshape([len(flattened_data), 1])
    res = fluid.LoDTensor()
    res.set(flattened_data, place)
    res.set_lod([lod])
    return res


def lodtensor_to_ndarray(lod_tensor):
    """conver lodtensor to ndarray
    """
    dims = lod_tensor.get_dims()
    ret = np.zeros(shape=dims).astype('float32')
    for i in xrange(np.product(dims)):
        ret.ravel()[i] = lod_tensor.get_float_element(i)
    return ret, lod_tensor.lod()