一、torch的安装
基于直接设备情况,选择合适的torch版本,有显卡的建议安装GPU版本,可以通过nvidia-smi
命令来查看显卡驱动的版本,在官网中根据cuda版本,选择合适的版本号,下面是安装示例代码
GPU:
pip install torch==2.5.0 torchvision==0.20.0 torchaudio==2.5.0 --index-url https://download.pytorch.org/whl/cu124
CPU:
pip install torch==2.5.0 torchvision==0.20.0 torchaudio==2.5.0 --index-url https://download.pytorch.org/whl/cpu
!!!,因为torch包很大,有3个G左右,下载时请选择稳定的网络环境,预防因网络波动导致安装失败。
二、tensor简述
PyTorch会将数据封装成张量(Tensor)进行计算,所谓张量就是元素为相同类型的多维矩阵。张量可以在 GPU 上加速运行。
2.1、概述
张量是一个多维数组,通俗来说可以看作是扩展了标量、向量、矩阵的更高维度的数组。张量的维度决定了它的形状(Shape),常见的张量有:
- 标量:0阶张量;
- 向量:1阶张量;
- 矩阵:2阶张量;
- 其他:例如高阶张量、图像、视频等复杂数据结构。
2.2、特点
- 动态计算图
- GPU支持
- 自动微分
2.3、数据类型
PyTorch中有3种数据类型:浮点数、整数、布尔。其中,浮点数和整数又分为8位、16位、32位、64位,加起来共9种。
三、tensor创建---- torch.tensor:根据指定的数据创建张量。
3.1、标量
def th_touchs():
# ? 创建一个一阶张量
t1 = torch.tensor(5)
print(t1.shape) # 打印张量的形状
print(t1.dtype) # 打印张量的数据类型
print(type(t1)) # 打印张量的类型
3.2、数组
# 数组
data = np.array([1, 2, 3, 4])
x = torch.tensor(data)
print(x)
3.3、列表
# 列表
data = [1,2,3,4]
x = torch.tensor(data)
print(x)
3.4、指定张量的数据类型
data = np.array([1, 2, 3, 4],dtype = int)
x = torch.tensor(data)
print(x)
3.5、指定张量的执行设备
data = np.array([1, 2, 3, 4],divice = "cuda")
x = torch.tensor(data)
print(x)
3.6、其他指定类型张量
"""
创建张量
"""
import torch
# ! 创建线性张量
def test_linear():
# 左闭右开区间,步长为2
x = torch.arange(1, 10, 2)
# print(x)
#闭区间,自动等差步长
y = torch.linspace(1, 10, 4)
# print(y)
# 等比数列,base表示指数,表示3^1~3^10取5个数
z = torch.logspace(1, 10, 10,base=2)
# print(z)
# ! 创建随机张量
def test_tensor():
# 设置随机数种子
torch.manual_seed(123)
# 获取当前随机数种子
# print(torch.initial_seed())
# 生成随机张量
x = torch.rand(10, 5)
# print(x)
# 生成正太分布的随机张量
y = torch.randn(2, 10)
# print(y)
# 自定义方差和均值
z = torch.normal(mean=10, std=2, size=(2, 2))
print(z)
# ! 创建0/1张量
def test_zeros():
# 创建指定形状的张量
x = torch.zeros((4, 4), dtype=torch.int64)
# print(x)
# 从数据中构建张量,传入的容器只能是tensor
x = torch.tensor([[1, 2, 3],
[4, 5, 6]])
x = torch.rand(3,3)
x = torch.zeros_like(x)
print(x)
def test_ones():
# 创建指定形状的张量
x = torch.ones((4, 4), dtype=torch.int64)
# print(x)
# 从数据中构建张量,传入的容器只能是tensor
x = torch.tensor([[1, 2, 3],
[4, 5, 6]])
x = torch.rand(3,3)
x = torch.ones_like(x)
print(x)
# ! 创建指定值张量
def test_full():
# 创建指定形状的张量
x = torch.full((4, 4), fill_value=7)
# print(x)
# 从数据中构建张量,传入的容器只能是tensor
x = torch.tensor([[1, 2, 3],
[4, 5, 6]])
x = torch.full_like(x, 11)
print(x)
# ! 创建单位矩阵张量
def test_eye():
# 创建指定形状的张量
x = torch.eye(4, 4)
print(x)
if __name__ == '__main__':
test_linear()
test_tensor()
test_zeros()
test_ones()
test_full()
test_eye()
tensor常用属性
"""
常见属性
"""
import torch
def th_tensor():
x = torch.tensor([1, 2, 3],device='cuda') # 指定创建到cuda/cpu的tensor
print(x.dtype) # 数据类型
print(x.device)# tensor所在的设备,默认是cpu
print(x.shape) # 形状
# 设备切换
def device_change():
# 方式1
# 将tensor创建在cuda设备上
x = torch.tensor([1, 2, 3], device='cuda')
# print(1,x.device)
# 方式2
# 先创建一个cpu上的tensor
x = torch.tensor([1, 2, 3])
# 将tensor移动到cuda设备上
x = x.to('cuda:0')
# print(2,x.device)
# 通过api获取设备是否有cuda
# 检查CUDA是否可用
res = torch.cuda.is_available()
print(res)
# 条件判断
c = 1 if 100>10 else 0
print(c)
# 根据CUDA可用性将tensor移动到cuda或cpu设备上
x.to("cuda" if torch.cuda.is_available() else "cpu")
print(x.device)
# 方式3
# 创建一个cpu上的tensor
x = torch.tensor([1, 2, 3])
print(x.device)
# 把tensor移动到cuda上
y = x.cuda() # 把tensor移动到cuda上
print(y.device)
# 根据CUDA可用性将tensor移动到cuda或cpu设备上
x = x.cuda() if torch.cuda.is_available() else x.cpu()
# 类型转换
def type_convert():
# 直接指定tensor的数据类型
x = torch.tensor([1, 2, 3], dtype=torch.float64)
print(x.dtype)
# 通过type方法转换tensor的数据类型
x = x.type(torch.int8)
print(x.dtype)
# 通过half方法转换tensor的数据类型
x = x.half()
print(x.dtype)
# 通过double方法转换tensor的数据类型
x = x.double()
print(x.dtype)
# 通过float方法转换tensor的数据类型
x = x.float()
print(x.dtype)
# 通过int方法转换tensor的数据类型
x = x.int()
print(x.dtype)
if __name__ == '__main__':
# print(th_tensor())
# device_change()
type_convert()
五、tensor数据类型转换
5.1、常规数据类型转换
import numpy as np
import torch
def th_data_np():
x = torch.tensor([1, 2, 3])
print(x)
# 把Tensor转换为numpy数组
x1 = x.numpy()
print(x1)
print(type(x1))
# x 和x1 是两个不同的对象,但它们都指向同一个数据存储空间
x1[0] = 100
print(x)
def th_data_copy():
x = torch.tensor([1, 2, 3])
print(x)
# 把Tensor转换为numpy数组并copy,copy()不会改变原来的数据
x1 = x.numpy().copy() # 注意这里是copy(深拷贝),还有一个view(浅拷贝)
print(x1)
print(type(x1))
x1[0] = 100
print(x)
print(x1)
def th_data_np_tensor():
# np转tensor,不共用内存(浅拷贝)
x = np.array([1, 2, 3])
print(x)
# 把numpy数组转换为Tensor
x1 = torch.tensor(x)
x[0] = 100
x1[0] = 200
print(x)
print(x1)
# from_numpy()会和原来的数组共享内存(深拷贝)
x2 = np.array([1, 2, 3])
x3 = torch.from_numpy(x2)
x2[0] = 100
x3[1] = 200
print(x2)
print(x3)
if __name__ == '__main__':
# th_data_np()
# th_data_copy()
th_data_np_tensor()
5.2、图片数据类型转换
from PIL import Image
import torch
from torchvision import transforms
def tensor_to_img():
pass
# ? 将图像转换为张量
def img_to_tensor():
path = './data/1.png'
img = Image.open(path)
print(img)
transfer = transforms.ToTensor()
img_tensor = transfer(img)
print(img_tensor.shape)
# ? 将张量转换为图像
def test():
# r = torch.rand(315,315)
# g = torch.rand(315,315)
# b = torch.rand(315,315)
img_tensor = torch.rand(4,315,315)
# print(img_tensor)
# print(img_tensor.shape)
# tensor转PIL对象
transfer = transforms.ToPILImage()
img = transfer(img_tensor)
img.show()
def test2():
prth = r"./data/1.png"
img = Image.open(prth)
print(img)
transfer = transforms.ToTensor()
img_tensor = transfer(img)
img_tensor[0] = 255
tensor2pil = transforms.ToPILImage()
img_pil = tensor2pil(img_tensor)
img_pil.show()
img_pil.save('./data/2.png')
if __name__ == "__main__":
# img_to_tensor(
# test()
test2()
六、tensor常见操作
6.1、获取元素
"""
从tensor中获取元素
"""
import torch
def th_items():
# 标量
x = torch.tensor(1)
print(x.item())
# 一阶
x = torch.tensor([100])
print(x.item())
# ! 如果输入的数据大于一个,会报错
x = torch.tensor([1, 2])
# print(x.item())
if __name__ == '__main__':
th_items()
6.2、元素值运算
"""
元素值运算
"""
import torch
def th_cction():
torch.manual_seed(666)
x = torch.randint(1, 10, (3, 3))
print(x)
# ? 加
x1 = x.add(100)
print(x1)
# !带_结尾的函数,基本上都是在原数据进行操作
x.add_(200)
print(x)
# ? 减
x2 = x.sub(50)
print(x2)
x.sub_(40)
print(x)
# ? 乘
x3 = x.mul(2)
print(x3)
x.mul_(2)
print(x)
# ? 除
x4 = x.div(2)
print(x4)
x.div_(2)
print(x)
# ? 幂
x5 = x.pow(2)
print(x5)
x.pow_(2)
print(x)
# ?
x6 = x**2
print(x6)
if __name__ == '__main__':
th_cction()
6.3、阿达玛积
"""
计算阿达玛积
"""
import torch
def adama():
x1 = torch.tensor([[1,2],
[3,4]])
x2 = torch.tensor([[1,2],
[3,4]])
#! Adama积矩阵形状必须相同
x3 = x1 * x2
print(x3)
if __name__ == '__main__':
adama()
6.4、相乘
"""
矩阵相乘
"""
import torch
def dot():
x1 = torch.tensor([[1,2],
[3,4]])
x2 = torch.tensor([[1,2],
[3,4]])
x3 = torch.matmul(x1, x2)
x3 = x1.matmul(x2)
x3 = x1 @ x2
x3 = x1.mm(x2)
print(x3)
if __name__ == '__main__':
dot()
6.5、索引
"""
索引
"""
import torch
def index():
# tensor 的布尔运算
# torch.manual_seed(66)
# x = torch.randint(0, 10, (5, 5))
# print(x)
# x1 = x > 7
# print(x1)
# x3 = x[x1]
# print(x3)
# print(x[x % 2 == 0])
# 创建一个5x5的tensor作为示例数据
x = torch.tensor([
[2, 3, 2000, 10, 20], # 满足条件:偶数,奇数,闰年
[1, 2, 2001, 30, 40], # 不满足条件:第一列是奇数
[4, 5, 2004, 50, 60], # 满足条件:偶数,奇数,闰年
[3, 7, 1900, 70, 80], # 不满足条件:第三列不是闰年(虽然能被4整除,但也能被100整除且不能被400整除)
[6, 9, 1600, 90, 100] # 满足条件:偶数,奇数,闰年
])
# 找出第一列是偶数,第二列是奇数,第三列是闰年的行中的第四列和第五列的数据
print(x[(x[:, 0] % 2 == 0) & (x[:, 1] % 2 != 0) & ((x[:, 2] % 4 == 0) | (x[:, 2] % 400 == 0))][:, [3, 4]])
# ? 索引赋值
def index_ass():
torch.manual_seed(66)
x = torch.randint(1,10,(5, 5))
print(x)
x1 = x[1,1]
print(x1)
x[1,1] = 100
print(x)
x[:,3] = 200
print(x)
x[:,:] = 300
print(x)
x.fill_(400)
print(x)
if __name__ == "__main__":
# index()
index_ass()
6.6、拼接
"""
拼接
"""
import torch
from PIL import Image
from torchvision import transforms
def montage_cat():
x = torch.randint(1, 10, (3,3))
y = torch.randint(1, 10, (3,3))
print(x)
print(y)
# cat():不会增加维度
# 拼接,dim=0:按行拼接,dim=1:按列拼接
z = torch.cat([x, y], dim=0)
print(z)
def montage_stack():
torch.manual_seed(66)
x = torch.randint(1, 10, (3,3))
y = torch.randint(1, 10, (3,3))
# print(x)
# print(y)
# ! stack():会增加维度(张量级别)
# ! 堆叠,dim=0:按行堆叠,dim=1:按列堆叠
# ! 要堆叠的张量必须具有相同的形状
z = torch.stack([x, y],dim=1)
print(z)
def montage_img():
# 加载本地图像为PIL
img = Image.open('./data/1.png')
transfer = transforms.ToTensor()
img_tensor = transfer(img)
# print(img_tensor)
# print(img_tensor.shape)
res = torch.stack([img_tensor[0], img_tensor[1],img_tensor[2]], dim=0)
if __name__ == '__main__':
# montage_cat()
# montage_stack()
montage_img()
6.7、形状
"""
形状操作
"""
import torch
def th_reshape():
x = torch.reshape(x, (-1, 5))
x1 = torch.reshape(x, (2, 6))
print(x)
x2 = torch.reshape(x, (2, 3, 2))
print(x2)
def th_view():
x = torch.randint(1, 10,(4, 3))
# print(x)
# !改变形状,由于没有改变原x中的内存空间,因此改变形状操作比reshape快
x1 = x.view((2, 6))
# print(x1)
# x2 = torch.randit(1, 10,(4, 3))
# # x3 = torch.reshape(x2, (2, 6))
# x4 = x2.t() # 转置矩阵
# print(x4)
# # !改变形状,在内存中不连续的数据不能通过view转换
# x5 = x4.view(2,6)
# print(x5)
# ! 改变形状后,是否共享数据内存
x6 = torch.randint(1, 10, (4, 3))
x7 = x6.view(2, 6)
x6[1,1] = 999
print(x6)
print(x7)
# ? 改变维度0
def rh_transpose():
x1 = torch.randint(1, 10, (4, 3))
print(x1,x1.shape)
# ! transpose(x, dim0, dim1), 交换两个维度
x2 = torch.transpose(x1, 0, 1)
print(x2,x2.shape)
# ? 改变维度1
def th_permute():
x1 = torch.randint(0, 255, (3, 512,360))
print(x1.shape)
x2 = x1.permute(1, 2, 0)
print(x2.shape)
# ? 改变为1维
def th_flatten():
x1 = torch.randint(0, 255, (4, 3))
# print(x1)
x2 = x1.flatten()
# print(x2)
x3 = torch.randint(0, 255, (3, 4, 2, 2))
print(x3)
x4 = x3.flatten(start_dim=1, end_dim=-1)
print(x4)
# ? 数据升维
def th_unsqueeze():
x1 = torch.randint(0, 255, (4, 3))
print(x1)
# ! 0,表示在0处插入一个维度
x2 = x1.unsqueeze(0)
print(x2.shape)
# ? 数据降维
def th_squeeze():
x1 = torch.randint(0, 255, (1, 4, 3, 1))
print(x1)
x2 = x1.squeeze()
print(x2)
x3 = x1.squeeze(0).squeeze(-1)
print(x3)
# ? 数据分割
def th_split():
x1 = torch.randint(0, 255, (4, 3))
print(x1)
# ! split(),表示每个tensor有2行
x2, x3 = torch.split(x1, 2)
# print(x2, x3)
# ! chunk(),表示将数据分割成多少份
x4 = torch.chunk(x1, 4)
print(x4)
# ? 广播
def th_broadcast():
a = torch.arange(1, 13).reshape(3, 4)
b = torch.arange(1, 5)
c = a + b
if __name__ == '__main__':
# th_reshape()
# th_view()
# rh_transpose()
# th_permute()
# th_flatten()
# th_unsqueeze()
# th_squeeze()
# th_split()
th_broadcast()
```