文章目录
1. torch.clamp
- 作用：将所有的值放到[min,max]之间
torch.clamp(input, min=None, max=None, *, out=None) → Tensor
f ( x ) = { m i n x < m i n x m i n < x < m a x m a x x > m a x (1) f(x)=\left\{ \begin{aligned} & min && x<min\\ & x&&min<x<max \\ & max && x>max \end{aligned} \right. \tag{1} f(x)=⎩⎪⎨⎪⎧minxmaxx<minmin<x<maxx>max(1)
- 代码
import torch
x = torch.randint(low=1,high=10,size=(3,4))
print(f'x={x}')
# 我们设置的最小值为4，最大值为 8
y = torch.clamp(x,4,8)
print(f'y={y}')
print(x == y)
- 结果
x=tensor([[8, 6, 2, 3],
[5, 5, 6, 1],
[8, 6, 3, 3]])
y=tensor([[8, 6, 4, 4],
[5, 5, 6, 4],
[8, 6, 4, 4]])
tensor([[ True, True, False, False],
[ True, True, True, False],
[ True, True, False, False]])
2. torchvision.transforms.ToTensor()
转换PIL图像或numpy.ndarray张量。此转换不支持torchscript.比如我们用opencv读取得到的数据为size=[height,width,channels],一般channels=3；0 <= height <= 255; 0 <=width<= 255;我们通过ToTensor;分两步；
- 将张量的通道数放在前面变成大小为[channels,height,width]
- 将张量里面的每个值都乘以255
- 代码
# -*- coding: utf-8 -*-
# @Project: zc
# @Author: zc
# @File name: torch.clamp_test
# @Create time: 2022/1/7 10:37
import cv2 as cv
import os
import torchvision
import torch
path = os.path.join(os.getcwd(), 'img', 'catdog.jpg')
img = cv.imread(path)
print(f'img.shape={img.shape}')
print(f'type(img)={type(img)}')
print(f'img={img}')
torchvision_tensor = torchvision.transforms.ToTensor()(img)
print(f'torchvision_tensor.shape={torchvision_tensor.shape}')
print(f'type(torchvision_tensor)={type(torchvision_tensor)}')
print(f'torchvision_tensor={torchvision_tensor}')
- 结果
img.shape=(825, 1045, 3)
type(img)=<class 'numpy.ndarray'>
img=[[[ 66 54 43]
[ 66 54 43]
[ 66 54 43]
...
[ 66 54 43]
[ 66 54 43]
[ 66 54 43]]
[[ 66 54 43]
[ 66 54 43]
[ 66 54 43]
...
[ 66 54 43]
[ 66 54 43]
[ 66 54 43]]
[[ 66 54 43]
[ 66 54 43]
[ 66 54 43]
...
[ 66 54 43]
[ 66 54 43]
[ 66 54 43]]
...
[[167 161 155]
[167 161 155]
[167 161 155]
...
[167 161 155]
[167 161 155]
[167 161 155]]
[[167 161 155]
[167 161 155]
[167 161 155]
...
[167 161 155]
[167 161 155]
[167 161 155]]
[[167 161 155]
[167 161 155]
[167 161 155]
...
[167 161 155]
[167 161 155]
[167 161 155]]]
torchvision_tensor.shape=torch.Size([3, 825, 1045])
type(torchvision_tensor)=<class 'torch.Tensor'>
torchvision_tensor=tensor([[[0.2588, 0.2588, 0.2588, ..., 0.2588, 0.2588, 0.2588],
[0.2588, 0.2588, 0.2588, ..., 0.2588, 0.2588, 0.2588],
[0.2588, 0.2588, 0.2588, ..., 0.2588, 0.2588, 0.2588],
...,
[0.6549, 0.6549, 0.6549, ..., 0.6549, 0.6549, 0.6549],
[0.6549, 0.6549, 0.6549, ..., 0.6549, 0.6549, 0.6549],
[0.6549, 0.6549, 0.6549, ..., 0.6549, 0.6549, 0.6549]],
[[0.2118, 0.2118, 0.2118, ..., 0.2118, 0.2118, 0.2118],
[0.2118, 0.2118, 0.2118, ..., 0.2118, 0.2118, 0.2118],
[0.2118, 0.2118, 0.2118, ..., 0.2118, 0.2118, 0.2118],
...,
[0.6314, 0.6314, 0.6314, ..., 0.6314, 0.6314, 0.6314],
[0.6314, 0.6314, 0.6314, ..., 0.6314, 0.6314, 0.6314],
[0.6314, 0.6314, 0.6314, ..., 0.6314, 0.6314, 0.6314]],
[[0.1686, 0.1686, 0.1686, ..., 0.1686, 0.1686, 0.1686],
[0.1686, 0.1686, 0.1686, ..., 0.1686, 0.1686, 0.1686],
[0.1686, 0.1686, 0.1686, ..., 0.1686, 0.1686, 0.1686],
...,
[0.6078, 0.6078, 0.6078, ..., 0.6078, 0.6078, 0.6078],
[0.6078, 0.6078, 0.6078, ..., 0.6078, 0.6078, 0.6078],
[0.6078, 0.6078, 0.6078, ..., 0.6078, 0.6078, 0.6078]]])
3. torchvision.transforms.Normalize
用均值和标准差对张量图像进行归一化。这个转换不支持PIL图像。给定n个通道的均值:([1]的均值，…，均值[n])和std:([1]的均值，…，std[n])，这个转换将标准化输入火炬的每个通道。*张量，即输出[通道]=(输入[通道]-平均[通道])/ std[通道]；即在给定一个均值和方差的时候，我们可以根据给定的均值和方差，将里面的每个通道里面的每个元素进行归一化处理。
o
u
t
p
u
t
_
c
h
a
n
n
e
l
=
i
n
p
u
t
_
c
h
a
n
n
e
l
−
m
e
a
n
(
c
h
a
n
n
e
l
)
s
t
d
(
c
h
a
n
n
e
l
)
(2)
output\_channel=\frac{input\_channel-mean(channel)}{std(channel)}\tag{2}
output_channel=std(channel)input_channel−mean(channel)(2)
- 代码
import torch
import torchvision
x = torch.arange(60, dtype=torch.float).reshape(3, 4, 5)
y = torchvision.transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[2, 2, 2])(x)
print(f'x={x}')
print(f'x_shape={x.shape}')
print(f'x[2,3,4]={x[2, 3, 4]}')
y_calculate = (x[2, 3, 4] - 0.5) / 2
print(f'y_calculate={y_calculate}')
print(f'y[2,3,4]={y[2, 3, 4]}')
print(f'y[2,3,4] = y_calculate is ={y[2, 3, 4] == y_calculate}')
print(f'y={y}')
print(f'y_shape={y.shape}')
- 分析
我们先生成一个张量，并把最后一个值 y[2,3,4]=59抽出来手动计算得到如下:
59 − 0.5 2 = 29.25 (3) \frac{59-0.5}{2}=29.25\tag{3} 259−0.5=29.25(3)
我们发现通过 Normalize 得到值跟手动计算的值一致。 - 结果
x=tensor([[[ 0., 1., 2., 3., 4.],
[ 5., 6., 7., 8., 9.],
[10., 11., 12., 13., 14.],
[15., 16., 17., 18., 19.]],
[[20., 21., 22., 23., 24.],
[25., 26., 27., 28., 29.],
[30., 31., 32., 33., 34.],
[35., 36., 37., 38., 39.]],
[[40., 41., 42., 43., 44.],
[45., 46., 47., 48., 49.],
[50., 51., 52., 53., 54.],
[55., 56., 57., 58., 59.]]])
x_shape=torch.Size([3, 4, 5])
x[2,3,4]=59.0
y_calculate=29.25
y[2,3,4]=29.25
y[2,3,4] = y_calculate is =True
y=tensor([[[-0.2500, 0.2500, 0.7500, 1.2500, 1.7500],
[ 2.2500, 2.7500, 3.2500, 3.7500, 4.2500],
[ 4.7500, 5.2500, 5.7500, 6.2500, 6.7500],
[ 7.2500, 7.7500, 8.2500, 8.7500, 9.2500]],
[[ 9.7500, 10.2500, 10.7500, 11.2500, 11.7500],
[12.2500, 12.7500, 13.2500, 13.7500, 14.2500],
[14.7500, 15.2500, 15.7500, 16.2500, 16.7500],
[17.2500, 17.7500, 18.2500, 18.7500, 19.2500]],
[[19.7500, 20.2500, 20.7500, 21.2500, 21.7500],
[22.2500, 22.7500, 23.2500, 23.7500, 24.2500],
[24.7500, 25.2500, 25.7500, 26.2500, 26.7500],
[27.2500, 27.7500, 28.2500, 28.7500, 29.2500]]])
y_shape=torch.Size([3, 4, 5])
4. torchvision.transforms.ToPILImage
将一个形状为C x H x W的张量tensor或形状H x W x C的numpy.ndarray到PIL图像，同时保持取值不变。
torchvision.transforms.ToPILImage(mode=None)
- 代码
# -*- coding: utf-8 -*-
# @Project: zc
# @Author: ZhangChu
# @File name: ToPILPicture_tets
# @Create time: 2022/1/7 17:49
# 1. 导入相关数据库
import torchvision
import os
import cv2 as cv
# 2.设置导入图片路径
path = os.path.join(os.getcwd(), 'img', 'cat2.jpg')
# 3.导入图片
input_img = cv.imread(path)
# 4. 查看导入图片的数据类型: numpy.ndarray
print(f'type(input_img)={type(input_img)}')
# 5. 将 numpy.ndarray 数据转换成 PIL图片格式
img_torchvision = torchvision.transforms.ToPILImage()(input_img)
print(f'type(img_torchvision)={type(img_torchvision)}')
# 6. 输出图片
img_torchvision.show()
- 结果
type(input_img)=<class 'numpy.ndarray'>
type(img_torchvision)=<class 'PIL.Image.Image'>
5. torchvision.transforms.Compose
在图像进行预处理的时候，我们经常的是需要将目标进行各种组合处理，为了满足上述要求，我们可以用Compose,将所有的预处理组合起来
class torchvision.transforms.Compose(transforms)
- 代码
Transforms_Compose = torchvision.transforms.Compose([
transforms.CenterCrop(10),
transforms.PILToTensor(),
transforms.ConvertImageDtype(torch.float)])
6. torch.optim.lr_scheduler.StepLR
torch.optim.lr_scheduler.StepLR(optimizer, step_size, gamma=0.1, last_epoch=- 1, verbose=False)
每一个步长时期，每个参数组的学习速率以伽马衰减。请注意，这种衰减可能与这个调度程序外部对学习速率的其他改变同时发生。当last_epoch=-1时，将初始lr设置为lr
- 我们可以根据此类来直接对优化器的学习率进行设置，一般是跟优化器一起使用的
- optimizer: 优化器 ：如 SGD,ADAM等等
7. torch.nn.DataParallel
在模块级实现数据并行
8. torch.utils.data.DataLoader
数据加载程序。结合一个数据集和一个采样器，并在给定数据集上提供一个可迭代对象
扫一扫
1万+
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
