这样做的一种方法是def my_odd_padding(list_of_2d_tensors, pad_value): # get the sizes of the matrices hs = [t_.shape[0] for t_ in list_of_2d_tensors] ws = [t_.shape[1] for t_ in list_of_2d_tensors] # allocate space for output result = torch.zeros(sum(hs), sum(ws)) result.add_(pad_value) fh = 0 fw = 0 for i, t_ in enumerate(list_of_2d_tensors): result[fh:fh+hs[i], fw:fw+ws[i]] = t_ fh += hs[i] fw += ws[i] return result 假设所有张量list_of_2d_tensors都相同dtype并且相同,device您可以result在使用分配时显式设置此 dtype 和设备torch.zeros