import torch
import numpy as np


def nantensor(*args, **kwargs):
    return torch.ones(*args, **kwargs) * np.nan


def nanmean(v, *args, inplace=False, **kwargs):
    if not inplace:
        v = v.clone()
    is_nan = torch.isnan(v)
    v[is_nan] = 0
    return v.sum(*args, **kwargs) / (~is_nan).float().sum(*args, **kwargs)