Pytorch：Tensor之间的拼接与拆分【拼接：cat、stack；拆分：split、chunk】

亦凉 2023-09-30 08:49 11阅读 0赞

![在这里插入图片描述][watermark_type_d3F5LXplbmhlaQ_shadow_50_text_Q1NETiBA5b-N6ICF44Gu5Lmx5aSq6YOO_size_20_color_FFFFFF_t_70_g_se_x_16_pic_center]

## 一、cat函数 ##

torch.catimport torch
    a=torch.randn(3,4) #随机生成一个shape（3，4）的tensort
    b=torch.randn(2,4) #随机生成一个shape（2，4）的tensor
    print("a:")
    print(a)
    print("b:")
    print(b)
    print("拼接结果：")
    #print(torch.cat([a,b],dim=1))#由于3！=2（a,b的行数不同故无法拼接）
    print(torch.cat([a,b],dim=0))#也就是原来的第0维维数增多
    #print(torch.cat([a,b],dim=0))返回一个shape（5，4）的tensor
    #把a和b拼接成一个shape（5，4）的tensor，
    #可理解为沿着行增加的方向（即纵向）拼接

a:
    tensor([[ 2.1754,  1.4698,  1.4103, -0.2498],
            [ 0.3248, -1.9372, -0.9310, -0.3833],
            [-0.3603, -0.0271, -0.1942, -0.0345]])
    b:
    tensor([[-0.3917,  0.1332, -1.0066,  0.3633],
            [-0.2378,  0.5224,  1.1371,  0.6401]])
    拼接结果：
    tensor([[ 2.1754,  1.4698,  1.4103, -0.2498],
            [ 0.3248, -1.9372, -0.9310, -0.3833],
            [-0.3603, -0.0271, -0.1942, -0.0345],
            [-0.3917,  0.1332, -1.0066,  0.3633],
            [-0.2378,  0.5224,  1.1371,  0.6401]])

## 二、stack函数 ##

创建一个新的维度。

要求：两个tensor拼接前“拼接维度”的形状要完全一致

torch.stack要求：两个tensor拼接前的形状完全一致import torch
    a=torch.randn(3,4)
    b=torch.randn(3,4)
    print("a: ")
    print(a)
    print("b: ")
    print(b)
    c=torch.stack([a,b],dim=0)
    #返回一个shape(2,3,4)的tensor,新增的维度2分别指向a和b
    # print("c: ")
    # print(c)
    # a:
    # tensor([[ 0.9722,  0.7518,  0.8787,  1.1068],
    #         [-0.3760, -0.3623, -0.9563,  0.3909],
    #         [ 0.7292, -0.0121, -0.4910,  2.1195]])
    # b:
    # tensor([[-0.4713, -2.5941,  1.8245,  0.2314],
    #         [ 1.3405,  0.3472,  1.1083,  0.7682],
    #         [-1.1995,  0.6853, -0.7180,  0.7114]])
    # c:
    # tensor([[[ 0.9722,  0.7518,  0.8787,  1.1068],
    #          [-0.3760, -0.3623, -0.9563,  0.3909],
    #          [ 0.7292, -0.0121, -0.4910,  2.1195]],
    #
    #         [[-0.4713, -2.5941,  1.8245,  0.2314],
    #          [ 1.3405,  0.3472,  1.1083,  0.7682],
    #          [-1.1995,  0.6853, -0.7180,  0.7114]]])
    d=torch.stack([a,b],dim=1)
    #返回一个shape（3,2,4）的tensor，新增的维度2分别指向相应的a的第i行和b的第i行
    print("d: ")
    print(d)
    # a:
    # tensor([[ 0.9923,  0.2121, -0.8024,  0.4230],
    #         [-0.6697, -2.7528, -1.2073,  0.9505],
    #         [ 0.5162, -0.9078, -0.6087, -0.4061]])
    # b:
    # tensor([[ 1.7505,  1.0785,  0.8404,  0.2812],
    #         [ 0.9416, -0.7041, -1.6120,  0.3687],
    #         [ 0.4658,  0.1827,  0.2341, -0.1813]])
    # d:
    # tensor([[[ 0.9923,  0.2121, -0.8024,  0.4230],
    #          [ 1.7505,  1.0785,  0.8404,  0.2812]],
    #
    #         [[-0.6697, -2.7528, -1.2073,  0.9505],
    #          [ 0.9416, -0.7041, -1.6120,  0.3687]],
    #
    #         [[ 0.5162, -0.9078, -0.6087, -0.4061],
    #          [ 0.4658,  0.1827,  0.2341, -0.1813]]])

这里的关键词参数dim的理解和cat方法中有些区别。

cat方法中可以理解为原tensor的维度，dim=0，就是沿着原来的0轴进行拼接，dim=1，就是沿着原来的1轴进行拼接。

stack方法中的dim则是指向新增维度的位置，dim=0，就是在新形成的tensor的维度的第0个位置新插入维度

## 三、split函数 ##

split是根据长度去拆分tensor

import torch
    a=torch.randn(3,4)
    print('a :')
    print(a)
    print("按维度0拆分 ： ")print(a.split([1,2],dim=0))
    
    #把维度0按照长度[1,2]拆分，形成2个tensor，
    #shape（1，4）和shape（2，4）
    # a :
    # tensor([[ 0.2626,  0.9178, -1.3622, -0.9441],
    #         [-0.1259, -0.3336,  0.2441, -0.2219],
    #         [ 1.5535,  0.7683, -1.7978, -1.1680]])
    # 按维度0拆分 ：
    # (tensor([[ 0.2626,  0.9178, -1.3622, -0.9441]]), tensor([[-0.1259, -0.3336,  0.2441, -0.2219],
    #         [ 1.5535,  0.7683, -1.7978, -1.1680]]))
    print("按维度1拆分 ： ")
    print(a.split([2,2],dim=1))
    #把维度1按照长度[2,2]拆分，形成2个tensor，
    #shape（3，2）和shape（3，2）
    # a :
    # tensor([[-0.5111,  1.3557, -0.1616, -0.2014],
    #         [-1.1011,  1.0982, -0.1794, -0.3510],
    #         [-1.4451, -2.1550, -1.9542, -1.6998]])
    # 按维度1拆分 ：
    # (tensor([[-0.5111,  1.3557],
    #         [-1.1011,  1.0982],
    #         [-1.4451, -2.1550]]), tensor([[-0.1616, -0.2014],
    #         [-0.1794, -0.3510],
    #         [-1.9542, -1.6998]]))

## 四、chunk函数 ##

chunk可以理解为均等分的split，但是当维度长度不能被等分份数整除时，虽然不会报错，但可能结果与预期的不一样，建议只在可以被整除的情况下运用

import torch
    a=torch.randn(4,6)
    print("a :")
    print(a)
    #print(a.chunk(2,dim=0))
    #返回2个shape（2，6）的tensor
    # a :
    # tensor([[ 0.8008,  0.6612, -0.0450,  0.3255, -0.4714, -0.2343],
    #         [ 1.3068, -0.2587,  1.4938,  0.1859,  0.4674,  0.0086],
    #         [ 0.4522,  0.0220, -0.2653, -0.0588,  1.1987,  0.7340],
    #         [ 0.1547, -0.2052, -0.8919, -0.8763, -0.6897,  0.2474]])
    # (tensor([[ 0.8008,  0.6612, -0.0450,  0.3255, -0.4714, -0.2343],
    #         [ 1.3068, -0.2587,  1.4938,  0.1859,  0.4674,  0.0086]]), tensor([[ 0.4522,  0.0220, -0.2653, -0.0588,  1.1987,  0.7340],
    #         [ 0.1547, -0.2052, -0.8919, -0.8763, -0.6897,  0.2474]]))
    print(a.chunk(2,dim=1))
    #返回2个shape（4，3）的tensor]]))
    a :
    tensor([[-0.4875,  1.4914,  0.2244, -0.5883, -0.5951, -0.4857],
            [-0.1344, -0.6973, -0.2042,  2.5817, -0.7972, -0.6522],
            [ 1.4379, -0.1185,  0.4457, -1.1168,  1.0184, -0.5088],
            [-0.7692,  1.4040, -0.2799,  1.1515,  0.2329,  0.4926]])
    (tensor([[-0.4875,  1.4914,  0.2244],
            [-0.1344, -0.6973, -0.2042],
            [ 1.4379, -0.1185,  0.4457],
            [-0.7692,  1.4040, -0.2799]]), tensor([[-0.5883, -0.5951, -0.4857],
            [ 2.5817, -0.7972, -0.6522],
            [-1.1168,  1.0184, -0.5088],
            [ 1.1515,  0.2329,  0.4926]]))

--------------------

参考资料：  
[Pytorch:Tensor的合并与分割][Pytorch_Tensor]  
[pytorch tensor 的拼接和拆分][pytorch tensor]

[watermark_type_d3F5LXplbmhlaQ_shadow_50_text_Q1NETiBA5b-N6ICF44Gu5Lmx5aSq6YOO_size_20_color_FFFFFF_t_70_g_se_x_16_pic_center]: https://img-blog.csdnimg.cn/709db175db9c470cbd5a67674d2357e7.png?x-oss-process=image/watermark,type_d3F5LXplbmhlaQ,shadow_50,text_Q1NETiBA5b-N6ICF44Gu5Lmx5aSq6YOO,size_20,color_FFFFFF,t_70,g_se,x_16#pic_center
[Pytorch_Tensor]: https://www.jianshu.com/p/4e57dbe1d281
[pytorch tensor]: http://t.zoukankan.com/tingtin-p-13579890.html