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
- tiup telemetry
- IDEA file encoding modification
- typeScript - Partially apply a function
- 2022杭电多校第三场 K题 Taxi
- Flask框架 根据源码分析可扩展点
- "No title"
- 软件测试面试题：您以往所从事的软件测试工作中，是否使用了一些工具来进行软件缺陷(Bug)的管理?如果有，请结合该工具描述软件缺陷(Bug)跟踪管理的流程？
- 【论文笔记】—低照度图像增强—Unsupervised—EnlightenGAN—2019-TIP
- 论文解读（ AF-GCL）《Augmentation-Free Graph Contrastive Learning with Performance Guarantee》
- 简单的顺序结构程序（C语言）
猜你喜欢
【idea】idea配置sql格式化
性能测试如何准备测试数据
简单的顺序结构程序（C语言）
Will domestic websites use Hong Kong servers be blocked?
redis可视化管理软件Redis Desktop Manager2022
子连接中的参数传递
leetcode: 266. All Palindromic Permutations
英特尔WiFi 7产品将于2024年亮相 最高速度可达5.8Gbps
matlab中rcosdesign函数升余弦滚降成型滤波器
The master teaches you the 3D real-time character production process, the game modeling process sharing
随机推荐
【LeetCode】双指针题解汇总
E - Distance Sequence （前缀和优化dp
动态上传jar包热部署
NMS原理及其代码实现
英特尔WiFi 7产品将于2024年亮相 最高速度可达5.8Gbps
2022牛客多校训练第二场 J题 Link with Arithmetic Progression
关于使用read table 语句
matlab中rcosdesign函数升余弦滚降成型滤波器
[230] Execute command error after connecting to Redis MISCONF Redis is configured to save RDB snapshots
STC89C52RC的P4口的应用问题
元宇宙：未来我们的每一个日常行为是否都能成为赚钱工具？
【LeetCode】图解 904. 水果成篮
翁恺C语言程序设计网课笔记合集
怎样进行在不改变主线程执行的时候,进行日志的记录
2022杭电多校 第三场 B题 Boss Rush
[230]连接Redis后执行命令错误 MISCONF Redis is configured to save RDB snapshots
D - I Hate Non-integer Number （选数的计数dp
2022 Hangzhou Electric Multi-School Training Session 3 1009 Package Delivery
2022牛客多校第三场 A Ancestor
软件测试面试题：软件测试类型都有哪些？