当前位置：网站首页>tensor.nozero(),面具,面具
tensor.nozero(),面具,面具
2022-08-05 00:23:00 【Le0v1n】
1. nozero()
https://pytorch.org/docs/stable/generated/torch.nonzero.html?highlight=nonzero#torch.nonzero
torch.nonzero(input, *, out=None, as_tuple=False) → LongTensor or tuple of LongTensors
torch.nonzero(..., as_tuple=False) (default)returns a2-Dtensor where each row is the index for a nonzero value.torch.nonzero(..., as_tuple=False) (默认)返回一个2-D张量,where each row is an index with a non-zero value.
tensor.nozero()默认返回一个2维的tensor, Inside is eligible indexes.
举个例子:
import torch
x = torch.randint(low=2, high=3, size=[2, 3])
idx = x.nonzero()
print(f"x:\n{
x}\n")
print(f"idx:\n{
idx}\n")
print(f"x.size: {
x.size()}")
print(f"idx.size: {
idx.size()}")
结果:
x:
tensor([[2, 2, 2],
[2, 2, 2]])
idx:
tensor([[0, 0],
[0, 1],
[0, 2],
[1, 0],
[1, 1],
[1, 2]])
x.size: torch.Size([2, 3])
idx.size: torch.Size([6, 2])
可以看到, 这里的idx是用2-D tensor来表示xthe index of the eligible element in.
因为
x有两个维度, 所以idx中的size()[1]也是2
如果输入是多维的tensor, Then it means that it will only be used2-D的tensor. 例子如下:
import torch
x = torch.randint(low=2, high=3, size=[2, 3, 4])
idx = x.nonzero()
print(f"x:\n{
x}\n")
print(f"idx:\n{
idx}\n")
print(f"x.size: {
x.size()}")
print(f"idx.size: {
idx.size()}")
结果:
x:
tensor([[[2, 2, 2, 2],
[2, 2, 2, 2],
[2, 2, 2, 2]],
[[2, 2, 2, 2],
[2, 2, 2, 2],
[2, 2, 2, 2]]])
idx:
tensor([[0, 0, 0],
[0, 0, 1],
[0, 0, 2],
[0, 0, 3],
[0, 1, 0],
[0, 1, 1],
[0, 1, 2],
[0, 1, 3],
[0, 2, 0],
[0, 2, 1],
[0, 2, 2],
[0, 2, 3],
[1, 0, 0],
[1, 0, 1],
[1, 0, 2],
[1, 0, 3],
[1, 1, 0],
[1, 1, 1],
[1, 1, 2],
[1, 1, 3],
[1, 2, 0],
[1, 2, 1],
[1, 2, 2],
[1, 2, 3]])
x.size: torch.Size([2, 3, 4])
idx.size: torch.Size([24, 3])
因为
x有3个维度, 所以idx中的size()[1]也是3
再举个例子:
import torch
x = torch.randint(low=2, high=3, size=[2, 3, 4, 2])
idx = x.nonzero()
print(f"x:\n{
x}\n")
print(f"idx:\n{
idx}\n")
print(f"x.size: {
x.size()}")
print(f"idx.size: {
idx.size()}")
结果:
x:
tensor([[[[2, 2],
[2, 2],
[2, 2],
[2, 2]],
[[2, 2],
[2, 2],
[2, 2],
[2, 2]],
[[2, 2],
[2, 2],
[2, 2],
[2, 2]]],
[[[2, 2],
[2, 2],
[2, 2],
[2, 2]],
[[2, 2],
[2, 2],
[2, 2],
[2, 2]],
[[2, 2],
[2, 2],
[2, 2],
[2, 2]]]])
idx:
tensor([[0, 0, 0, 0],
[0, 0, 0, 1],
[0, 0, 1, 0],
[0, 0, 1, 1],
[0, 0, 2, 0],
[0, 0, 2, 1],
[0, 0, 3, 0],
[0, 0, 3, 1],
[0, 1, 0, 0],
[0, 1, 0, 1],
[0, 1, 1, 0],
[0, 1, 1, 1],
[0, 1, 2, 0],
[0, 1, 2, 1],
[0, 1, 3, 0],
[0, 1, 3, 1],
[0, 2, 0, 0],
[0, 2, 0, 1],
[0, 2, 1, 0],
[0, 2, 1, 1],
[0, 2, 2, 0],
[0, 2, 2, 1],
[0, 2, 3, 0],
[0, 2, 3, 1],
[1, 0, 0, 0],
[1, 0, 0, 1],
[1, 0, 1, 0],
[1, 0, 1, 1],
[1, 0, 2, 0],
[1, 0, 2, 1],
[1, 0, 3, 0],
[1, 0, 3, 1],
[1, 1, 0, 0],
[1, 1, 0, 1],
[1, 1, 1, 0],
[1, 1, 1, 1],
[1, 1, 2, 0],
[1, 1, 2, 1],
[1, 1, 3, 0],
[1, 1, 3, 1],
[1, 2, 0, 0],
[1, 2, 0, 1],
[1, 2, 1, 0],
[1, 2, 1, 1],
[1, 2, 2, 0],
[1, 2, 2, 1],
[1, 2, 3, 0],
[1, 2, 3, 1]])
x.size: torch.Size([2, 3, 4, 2])
idx.size: torch.Size([48, 4])
因为
x有4个维度, 所以idx中的size()[1]也是4
终于, We can find the conclusion:
tensor.nozero()默认返回的是一个2-Dtensor, Rows represent indices of non-zero elements, 列的大小 = 输入tensor维度的数量
2. mask
in handYOLOv3时, 需要写mask, Generally, it is judged whether it is a positive sample according to whether the confidence level is greater than the threshold.:
- ≥ thresh: 张样本
- < thresh: 负样本
在YOLOv3中, The forward reasoning results are generally of the network [N, H, W, 3, 8], 其中:
N: batch sizeH: heightW: weight3: 3prediction scale8: conf + loc + cls = 1 + 4 + 3
To get confidenceconf, 因此output[..., 0], 由此就可以得到mask:
output = net(input) # [N, H, W, 3, 8]
mask = output[..., 0] > thresh # [N, H, W, 3]
那么问题来了: 为什么得到的mask的size为[N, H, W, 3]呢?
我们讲一下maskThis method is doing.
import torch
x = torch.randint(low=2, high=3, size=[2, 3])
mask = x[..., 0] > 1 # The first element in the last dimension if greater than1, then the previous dimension is returnedTrue, 否则返回False
print(f"x:\n{
x}\n")
print(f"x[..., 0]:\n{
x[..., 0]}\n")
print(f"x[..., 0].size:\n{
x[..., 0].size()}\n")
print(f"mask:\n{
mask}\n")
print(f"mask.size:\n{
mask.size()}\n")
x:
tensor([[2, 2, 2],
[2, 2, 2]])
x[..., 0]:
tensor([2, 2])
x[..., 0].size:
torch.Size([2])
mask:
tensor([True, True])
mask.size:
torch.Size([2])
Maybe it's not obvious enough, 那么我们将x的值修改一下:
import torch
x = torch.tensor(data=[[0, 2, 2],
[0, 2, 2]], dtype=torch.int8)
mask = x[..., 0] > 1 # The first element in the last dimension if greater than1, then the previous dimension is returnedTrue, 否则返回False
print(f"x:\n{
x}\n")
print(f"x[..., 0]:\n{
x[..., 0]}\n")
print(f"x[..., 0].size:\n{
x[..., 0].size()}\n")
print(f"mask:\n{
mask}\n")
print(f"mask.size:\n{
mask.size()}\n")
结果:
x:
tensor([[0, 2, 2],
[0, 2, 2]], dtype=torch.int8)
x[..., 0]:
tensor([0, 0], dtype=torch.int8)
x[..., 0].size:
torch.Size([2])
mask:
tensor([False, False])
mask.size:
torch.Size([2])
看起来, This function takes the first column(The description here is not accurate). 那么我们在将xExpand the dimension of:
import torch
x = torch.randint(low=0, high=3, size=[2, 3, 2])
mask = x[..., 0] > 1 # The first element in the last dimension if greater than1, then the previous dimension is returnedTrue, 否则返回False
print(f"x:\n{
x}\n")
print(f"x[..., 0]:\n{
x[..., 0]}\n")
print(f"x[..., 0].size:\n{
x[..., 0].size()}\n")
print(f"mask:\n{
mask}\n")
print(f"mask.size:\n{
mask.size()}\n")
结果:
x:
tensor([[[2, 1],
[2, 1],
[1, 0]],
[[2, 0],
[1, 2],
[0, 0]]])
x[..., 0]:
tensor([[2, 2, 1],
[2, 1, 0]])
x[..., 0].size:
torch.Size([2, 3])
mask:
tensor([[ True, True, False],
[ True, False, False]])
mask.size:
torch.Size([2, 3])
The conclusion is still true, 我们再将x扩充至 [N, H, W, 3]:
import torch
x = torch.randint(low=0, high=3, size=[2, 3, 2, 1])
mask = x[..., 0] > 1 # The first element in the last dimension if greater than1, then the previous dimension is returnedTrue, 否则返回False
print(f"x:\n{
x}\n")
print(f"x[..., 0]:\n{
x[..., 0]}\n")
print(f"x[..., 0].size:\n{
x[..., 0].size()}\n")
print(f"mask:\n{
mask}\n")
print(f"mask.size:\n{
mask.size()}\n")
x:
tensor([[[[0],
[0]],
[[0],
[1]],
[[2],
[2]]],
[[[1],
[2]],
[[2],
[2]],
[[2],
[2]]]])
x[..., 0]:
tensor([[[0, 0],
[0, 1],
[2, 2]],
[[1, 2],
[2, 2],
[2, 2]]])
x[..., 0].size:
torch.Size([2, 3, 2])
mask:
tensor([[[False, False],
[False, False],
[ True, True]],
[[False, True],
[ True, True],
[ True, True]]])
mask.size:
torch.Size([2, 3, 2])
结论是对的.
3. [mask]
上面我们分析了maskwhat are you taking, 那么将mask应用到x中(x[mask])会是怎么样的呢?
import torch
x = torch.randint(low=0, high=3, size=[2, 3])
mask = x[..., 0] > 1 # The first element in the last dimension if greater than1, then the previous dimension is returnedTrue, 否则返回False
filtered_x = x[mask]
print(f"x:\n{
x}\n")
print(f"x.size:\n{
x.size()}\n")
print("-" * 50)
print(f"x[..., 0]:\n{
x[..., 0]}\n")
print(f"x[..., 0].size:\n{
x[..., 0].size()}\n")
print("-" * 50)
print(f"mask:\n{
mask}\n")
print(f"mask.size:\n{
mask.size()}\n")
print("-" * 50)
print(f"filtered_x:\n{
filtered_x}\n")
print(f"filtered_x.size:\n{
filtered_x.size()}\n")
结果2:
x:
tensor([[2, 2, 1],
[0, 0, 2]])
x.size:
torch.Size([2, 3])
--------------------------------------------------
x[..., 0]:
tensor([2, 0])
x[..., 0].size:
torch.Size([2])
--------------------------------------------------
mask:
tensor([ True, False])
mask.size:
torch.Size([2])
--------------------------------------------------
filtered_x:
tensor([[2, 2, 1]])
filtered_x.size:
torch.Size([1, 3])
mask负责筛选"列", x[mask]Responsible for fetching eligible rows, and this line is a2D tensor.
结果2:
x:
tensor([[0, 1, 0],
[1, 0, 0]])
x.size:
torch.Size([2, 3])
--------------------------------------------------
x[..., 0]:
tensor([0, 1])
x[..., 0].size:
torch.Size([2])
--------------------------------------------------
mask:
tensor([False, False])
mask.size:
torch.Size([2])
--------------------------------------------------
filtered_x:
tensor([], size=(0, 3), dtype=torch.int64)
filtered_x.size:
torch.Size([0, 3])
mask负责筛选"列", x[mask]Responsible for fetching eligible rows, and this line is a2D tensor.
我们将xdimensional improvement:
import torch
x = torch.randint(low=0, high=3, size=[2, 3, 2])
mask = x[..., 0] > 1 # The first element in the last dimension if greater than1, then the previous dimension is returnedTrue, 否则返回False
filtered_x = x[mask]
print(f"x:\n{
x}\n")
print(f"x.size:\n{
x.size()}\n")
print("-" * 50)
print(f"x[..., 0]:\n{
x[..., 0]}\n")
print(f"x[..., 0].size:\n{
x[..., 0].size()}\n")
print("-" * 50)
print(f"mask:\n{
mask}\n")
print(f"mask.size:\n{
mask.size()}\n")
print("-" * 50)
print(f"filtered_x:\n{
filtered_x}\n")
print(f"filtered_x.size:\n{
filtered_x.size()}\n")
结果1:
x:
tensor([[[0, 2],
[0, 2],
[0, 2]],
[[2, 2],
[1, 0],
[1, 2]]])
x.size:
torch.Size([2, 3, 2])
--------------------------------------------------
x[..., 0]:
tensor([[0, 0, 0],
[2, 1, 1]])
x[..., 0].size:
torch.Size([2, 3])
--------------------------------------------------
mask:
tensor([[False, False, False],
[ True, False, False]])
mask.size:
torch.Size([2, 3])
--------------------------------------------------
filtered_x:
tensor([[2, 2]])
filtered_x.size:
torch.Size([1, 2])
结果2:
x:
tensor([[[1, 0],
[2, 0],
[2, 1]],
[[2, 0],
[2, 0],
[2, 1]]])
x.size:
torch.Size([2, 3, 2])
--------------------------------------------------
x[..., 0]:
tensor([[1, 2, 2],
[2, 2, 2]])
x[..., 0].size:
torch.Size([2, 3])
--------------------------------------------------
mask:
tensor([[False, True, True],
[ True, True, True]])
mask.size:
torch.Size([2, 3])
--------------------------------------------------
filtered_x:
tensor([[2, 0],
[2, 1],
[2, 0],
[2, 0],
[2, 1]])
filtered_x.size:
torch.Size([5, 2])
我们可以发现, 结论 mask负责筛选"列", x[mask]Responsible for fetching eligible rows, and this line is a2D tensor 依然是符合的, Let's expand the dimensions:
import torch
x = torch.randint(low=0, high=3, size=[2, 3, 2, 2])
mask = x[..., 0] > 1 # The first element in the last dimension if greater than1, then the previous dimension is returnedTrue, 否则返回False
filtered_x = x[mask]
print(f"x:\n{
x}\n")
print(f"x.size:\n{
x.size()}\n")
print("-" * 50)
print(f"x[..., 0]:\n{
x[..., 0]}\n")
print(f"x[..., 0].size:\n{
x[..., 0].size()}\n")
print("-" * 50)
print(f"mask:\n{
mask}\n")
print(f"mask.size:\n{
mask.size()}\n")
print("-" * 50)
print(f"filtered_x:\n{
filtered_x}\n")
print(f"filtered_x.size:\n{
filtered_x.size()}\n")
结果1:
x:
tensor([[[[0, 0],
[0, 1]],
[[1, 1],
[0, 0]],
[[0, 1],
[2, 1]]],
[[[2, 2],
[0, 1]],
[[0, 2],
[0, 2]],
[[0, 0],
[2, 2]]]])
x.size:
torch.Size([2, 3, 2, 2])
--------------------------------------------------
x[..., 0]:
tensor([[[0, 0],
[1, 0],
[0, 2]],
[[2, 0],
[0, 0],
[0, 2]]])
x[..., 0].size:
torch.Size([2, 3, 2])
--------------------------------------------------
mask:
tensor([[[False, False],
[False, False],
[False, True]],
[[ True, False],
[False, False],
[False, True]]])
mask.size:
torch.Size([2, 3, 2])
--------------------------------------------------
filtered_x:
tensor([[2, 1],
[2, 2],
[2, 2]])
filtered_x.size:
torch.Size([3, 2])
结果2:
x:
tensor([[[[2, 2],
[2, 2]],
[[1, 0],
[0, 2]],
[[0, 2],
[2, 2]]],
[[[0, 0],
[1, 0]],
[[1, 1],
[0, 1]],
[[2, 1],
[1, 0]]]])
x.size:
torch.Size([2, 3, 2, 2])
--------------------------------------------------
x[..., 0]:
tensor([[[2, 2],
[1, 0],
[0, 2]],
[[0, 1],
[1, 0],
[2, 1]]])
x[..., 0].size:
torch.Size([2, 3, 2])
--------------------------------------------------
mask:
tensor([[[ True, True],
[False, False],
[False, True]],
[[False, False],
[False, False],
[ True, False]]])
mask.size:
torch.Size([2, 3, 2])
--------------------------------------------------
filtered_x:
tensor([[2, 2],
[2, 2],
[2, 2],
[2, 1]])
filtered_x.size:
torch.Size([4, 2])
The conclusion remains!
4. 结论
Now we finally understand,
output = net(input) # [N, H, W, 3, 8]
mask = output[..., 0] > thresh # [N, H, W, 3]
mask: Take out the columns that meet the conditions (The dimension used for judgment disappears, 所以[N, H, W, 3, 8]->[N, H, W, 3])output[mask]: 取出符合条件的行 (是一个2-Dtensor)
举个例子:
def get_idx_and_info(self, output, thresh):
output = output.permute(0, 2, 3, 1) # [N, 24, H, W] -> [N, H, W, 3*8]
N, H, W, _ = output.size()
output = output.reshape(N, H, W, 3, -1) # N, H, W, 24] -> N, H, W, 3, 8]
# 定义掩码(符合条件的行)
mask = output[..., 0] > thresh # 置信度 > thresh # [N, H, W, 3] (It should be automaticsqueeze了)
idx = mask.nonzero() # size -> [Number of eligible numbers, 4] # 因为maskThe number of dimensions has4个, 所以这里是4,
info = output[mask] # size -> [Number of eligible numbers, 8] # 8 = c + loc + cls
return idx, info
边栏推荐
- SV 类的虚方法 多态
- Software test interview questions: BIOS, Fat, IDE, Sata, SCSI, Ntfs windows NT?
- 软件测试面试题：BIOS, Fat, IDE, Sata, SCSI, Ntfs windows NT？
- Flask框架 根据源码分析可扩展点
- [LeetCode] Summary of Matrix Simulation Related Topics
- 日志（logging模块）
- 2022牛客多校训练第二场 L题 Link with Level Editor I
- 机器学习(公式推导与代码实现)--sklearn机器学习库
- 2022 Niu Ke Summer Multi-School Training Camp 5 (BCDFGHK)
- TinyMCE disable escape
猜你喜欢
How to automatically push my new articles to my fans (very simple, can't learn to hit me)
leetcode: 266. All Palindromic Permutations
看图识字，DELL SC4020 / SCv2000 控制器更换过程
【云原生--Kubernetes】Pod控制器
matlab中rcosdesign函数升余弦滚降成型滤波器
2022 Niu Ke Summer Multi-School Training Camp 5 (BCDFGHK)
英特尔WiFi 7产品将于2024年亮相 最高速度可达5.8Gbps
【论文笔记】—低照度图像增强—Unsupervised—EnlightenGAN—2019-TIP
找不到DiscoveryClient类型的Bean
2 用D435i运行VINS-fusion
随机推荐
E - Many Operations (bitwise consideration + dp thought to record the result after the operation
僵尸进程和孤儿进程
2022 Hangzhou Electric Power Multi-School Session 3 Question B Boss Rush
10 种常见的BUG分类
ansible学习笔记分享-含剧本示例
英特尔WiFi 7产品将于2024年亮相 最高速度可达5.8Gbps
数据类型及输入输出初探（C语言）
【idea】idea配置sql格式化
Huggingface入门篇 II (QA)
redis可视化管理软件Redis Desktop Manager2022
[Cloud Native--Kubernetes] Pod Controller
what？测试/开发程序员要被淘汰了？年龄40被砍到了32？一瞬间，有点缓不过神来......
canvas 高斯模糊效果
找不到DiscoveryClient类型的Bean
Raw and scan of gorm
Mysql based
Software testing interview questions: test life cycle, the test process is divided into several stages, and the meaning of each stage and the method used?
电子行业MES管理系统的主要功能与用途
【LeetCode】Summary of Two Pointer Problems
"WEB Security Penetration Testing" (28) Burp Collaborator-dnslog out-band technology