Pytorch：Tensor的统计属性【norm：范数】、【mean、sum】【prod：累乘】【argmin、argmax：最值索引】【keepdim】【kthvalue：第k小的元素、topk】

￡神魔★判官ぃ 2023-09-30 08:49 11阅读 0赞

范数1：所有元素的绝对值的求和  
范数2：所有元素的绝对值的平方和的开方  
max（）：矩阵中的最大值以及相应的index  
min（）：矩阵中的最小值以及相应的index  
mean（）：平均值  
prod（）：累乘  
sum（）：求和  
argmax（）：返回最大值的index  
argmin（）：返回最小值的index  
argmax与argmin不带参数的话，会将矩阵先打平之后再寻找最值的index，这样找到的index一定只是一维的，而不是打平之前的index  
argmax（dim=1）来指定进行比较的方向

## 一、norm：范数 ##

范数1：所有元素的绝对值的求和

范数2：所有元素的绝对值的平方和的开方

a = torch.full([8],1)
    b = a.view(2,4)
    c = a.view(2,2,2)
    
    a.norm(1),b.norm(1),c.norm(1)
    #都是tensor(8)
    
    a.norm(2),b.norm(2),c.norm(2)
    #都是tensor(2.8284)  根号8

在指定的维数上面进行norm的查看，取哪个维度的范数，该维度就会被消掉。

a = torch.full([8],1)
    b = a.view(2,4)
    c = a.view(2,2,2)
    
    b.norm(1,dim=1)
    #tensor([4,4])  一共两行 按行取

![在这里插入图片描述][watermark_type_d3F5LXplbmhlaQ_shadow_50_text_Q1NETiBA5b-N6ICF44Gu5Lmx5aSq6YOO_size_17_color_FFFFFF_t_70_g_se_x_16]

## 二、mean、sum、prod、max、min、argmin、argmax ##

argmin、argmax：不加参数时，将矩阵打平为一维，然后取最小值、最大值的索引；  
![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]  
![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 1]

import torch
    
    a = torch.randn(4, 6)
    
    print('a.shape = ', a.shape)
    print('a = ', a)
    
    a_argmax = a.argmax()
    a_argmax_dim_0 = a.argmax(dim=0)
    a_argmax_dim_1 = a.argmax(dim=1)
    
    print('\na_argmax = ', a_argmax)
    print('a_argmax_dim_0 = ', a_argmax_dim_0)
    print('a_argmax_dim_1 = ', a_argmax_dim_1)

打印结果：

a.shape =  torch.Size([4, 6])
    a =  tensor([[ 0.4317, -0.9791, -1.8569,  0.9546,  1.5674, -0.7068],
      		      [ 0.4386, -1.5702,  0.9274,  0.9188, -1.3072, -1.2542],
       		     [ 0.7009, -0.1240,  1.0909,  0.5051,  1.4061, -0.7690],
       		     [-0.7982,  0.3734,  0.9044,  0.8111, -0.1462,  0.0558]])
    
    a_argmax =  tensor(4)
    a_argmax_dim_0 =  tensor([2, 3, 2, 0, 0, 3])
    a_argmax_dim_1 =  tensor([4, 2, 4, 2])

## 三、keepdim ##

![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 2]

## 四、topk、kthvalue ##

### 1、topk：前k个最值 ###

import torch
    
    a = torch.randn(4, 6)
    
    print('a.shape = ', a.shape)
    print('a = ', a)
    
    a_topk = a.topk(3, dim=1)
    print('\na_topk = ', a_topk)

打印结果：

a.shape =  torch.Size([4, 6])
    a =  tensor([[ 0.5713, -0.8378, -0.9350,  1.4258,  1.0426, -1.3174],
          		  [-1.5093, -0.9687,  1.1797,  0.9376, -0.2584,  0.3872],
        		    [ 1.7028, -1.4444, -0.5496,  0.0120, -0.2293, -1.5676],
       		     [-0.0805,  1.0818, -1.0642, -1.1229,  0.1533, -0.9521]])
    
    a_topk =  torch.return_types.topk(
    values=tensor([[ 1.4258,  1.0426,  0.5713],
      		      [ 1.1797,  0.9376,  0.3872],
      		      [ 1.7028,  0.0120, -0.2293],
      		      [ 1.0818,  0.1533, -0.0805]]),
    indices=tensor([[3, 4, 0],
        		    [2, 3, 5],
        		    [0, 3, 4],
        		    [1, 4, 0]]))
    
    Process finished with exit code 0

### 2、kthvalue：排序后第k小的元素的值 ###

import torch
    
    a = torch.randn(4, 6)
    
    print('a.shape = ', a.shape)
    print('a = ', a)
    
    a_kthvalue = a.kthvalue(3, dim=1)
    print('\na_kthvalue = ', a_kthvalue)

打印结果：

a.shape =  torch.Size([4, 6])
    a =  tensor([[-0.6937, -2.6252, -0.0817,  0.1022, -0.4947, -0.0497],
        	    [ 1.5205,  0.1827, -1.1397,  2.1021,  1.0715,  1.2167],
        	    [-1.6002,  1.1374, -1.2099,  2.4795, -0.6292, -1.1339],
        	    [ 0.9659, -0.1593, -0.0963,  0.8514, -1.0020,  0.9966]])
    
    a_kthvalue =  torch.return_types.kthvalue(
    								values=tensor([-0.4947,  1.0715, -1.1339, -0.0963]),
    								indices=tensor([4, 4, 5, 2])
    								)
    
    Process finished with exit code 0

## 五、compare：>、>=、<、<=、!=、== ##

![在这里插入图片描述][watermark_type_d3F5LXplbmhlaQ_shadow_50_text_Q1NETiBA5b-N6ICF44Gu5Lmx5aSq6YOO_size_17_color_FFFFFF_t_70_g_se_x_16 1]

### 1、> ###

import torch
    
    a = torch.randn(4, 6)
    
    print('a.shape = ', a.shape)
    print('a = ', a)
    
    a_0 = a > 0
    print('\na_0 = ', a_0)

打印结果：

a.shape =  torch.Size([4, 6])
    a =  tensor([[-1.3893,  1.4089, -0.1790, -0.4641, -2.9107, -1.2560],
          		  [ 0.1702,  0.3689, -1.4860,  0.5642,  0.8496,  0.4877],
        		    [ 0.7780, -0.5726,  1.0474,  0.7246,  0.2890, -1.1601],
        		    [ 1.2610,  0.1574, -0.7690,  0.7012,  0.2023,  0.8884]])
    
    a_0 =  tensor([[False,  True, False, False, False, False],
          		  [ True,  True, False,  True,  True,  True],
          		  [ True, False,  True,  True,  True, False],
         		   [ True,  True, False,  True,  True,  True]])
    
    Process finished with exit code 0

### 2、!= ###

import torch
    
    a = torch.randn(4, 6)
    
    print('a.shape = ', a.shape)
    print('a = ', a)
    
    a_0 = a != 0
    print('\na_0 = ', a_0)

打印结果：

a.shape =  torch.Size([4, 6])
    a =  tensor([[ 0.0204,  0.3504,  1.2778, -0.8016, -2.2371, -2.1239],
       		     [-0.3999,  2.6354,  3.4434,  0.0612, -0.7063, -0.2649],
        		    [ 0.1774, -0.1433, -1.0806, -2.4606, -0.4880,  0.4409],
        		    [-1.2463,  0.8048,  0.9639, -0.1631,  0.4157, -0.1088]])
    
    a_0 =  tensor([[True, True, True, True, True, True],
        		    [True, True, True, True, True, True],
         		   [True, True, True, True, True, True],
         		   [True, True, True, True, True, True]])
    
    Process finished with exit code 0

### 3、== ###

import torch
    
    a = torch.ones(4, 6)
    b = torch.randn(4, 6)
    
    a_b_eq01 = torch.eq(a, b)
    a_b_eq02 = a == b
    
    print('a_b_eq01 = ', a_b_eq01)
    print('a_b_eq02 = ', a_b_eq02)
    
    a_a_eq01 = torch.eq(a, a)
    a_a_eq02 = a == a
    
    print('\na_a_eq01 = ', a_a_eq01)
    print('a_a_eq02 = ', a_a_eq02)

打印结果：

a_b_eq01 =  tensor([[False, False, False, False, False, False],
            [False, False, False, False, False, False],
            [False, False, False, False, False, False],
            [False, False, False, False, False, False]])
    a_b_eq02 =  tensor([[False, False, False, False, False, False],
            [False, False, False, False, False, False],
            [False, False, False, False, False, False],
            [False, False, False, False, False, False]])
    
    a_a_eq01 =  tensor([[True, True, True, True, True, True],
            [True, True, True, True, True, True],
            [True, True, True, True, True, True],
            [True, True, True, True, True, True]])
    a_a_eq02 =  tensor([[True, True, True, True, True, True],
            [True, True, True, True, True, True],
            [True, True, True, True, True, True],
            [True, True, True, True, True, True]])
    
    Process finished with exit code 0

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

参考资料：  
[tensor的统计属性][tensor]

[watermark_type_d3F5LXplbmhlaQ_shadow_50_text_Q1NETiBA5b-N6ICF44Gu5Lmx5aSq6YOO_size_17_color_FFFFFF_t_70_g_se_x_16]: https://img-blog.csdnimg.cn/70ec76f731cf419da71ef148eae73c17.png?x-oss-process=image/watermark,type_d3F5LXplbmhlaQ,shadow_50,text_Q1NETiBA5b-N6ICF44Gu5Lmx5aSq6YOO,size_17,color_FFFFFF,t_70,g_se,x_16
[watermark_type_d3F5LXplbmhlaQ_shadow_50_text_Q1NETiBA5b-N6ICF44Gu5Lmx5aSq6YOO_size_20_color_FFFFFF_t_70_g_se_x_16]: https://img-blog.csdnimg.cn/6a91dd6364a047e58b6754c0deefaf0d.png?x-oss-process=image/watermark,type_d3F5LXplbmhlaQ,shadow_50,text_Q1NETiBA5b-N6ICF44Gu5Lmx5aSq6YOO,size_20,color_FFFFFF,t_70,g_se,x_16
[watermark_type_d3F5LXplbmhlaQ_shadow_50_text_Q1NETiBA5b-N6ICF44Gu5Lmx5aSq6YOO_size_20_color_FFFFFF_t_70_g_se_x_16 1]: https://img-blog.csdnimg.cn/d598df1b0a02444693c3b4c92d59ff6d.png?x-oss-process=image/watermark,type_d3F5LXplbmhlaQ,shadow_50,text_Q1NETiBA5b-N6ICF44Gu5Lmx5aSq6YOO,size_20,color_FFFFFF,t_70,g_se,x_16
[watermark_type_d3F5LXplbmhlaQ_shadow_50_text_Q1NETiBA5b-N6ICF44Gu5Lmx5aSq6YOO_size_20_color_FFFFFF_t_70_g_se_x_16 2]: https://img-blog.csdnimg.cn/5f65e9ba68ec403983013e36dddbc6dc.png?x-oss-process=image/watermark,type_d3F5LXplbmhlaQ,shadow_50,text_Q1NETiBA5b-N6ICF44Gu5Lmx5aSq6YOO,size_20,color_FFFFFF,t_70,g_se,x_16
[watermark_type_d3F5LXplbmhlaQ_shadow_50_text_Q1NETiBA5b-N6ICF44Gu5Lmx5aSq6YOO_size_17_color_FFFFFF_t_70_g_se_x_16 1]: https://img-blog.csdnimg.cn/142226150e2c4ab482e3258341840d3d.png?x-oss-process=image/watermark,type_d3F5LXplbmhlaQ,shadow_50,text_Q1NETiBA5b-N6ICF44Gu5Lmx5aSq6YOO,size_17,color_FFFFFF,t_70,g_se,x_16
[tensor]: https://www.cnblogs.com/Jason66661010/p/13602392.html