Pytorch笔记 1 Tensor和Variable

Tensor

首先导入有关的库

import torch
import numpy as np

numpy与tensor的转换

由numpy转换成tensor

对于numpy到tensor的转换，一般有两种方法

利用torch.Tensor

tensor_from_np1 = torch.Tensor(numpy_tensor)
tensor_from_np2 = torch.from_numpy(numpy_tensor)

下面是将tensor转化成numpy

numpy_array = tensor_from_np1.numpy()

如果此时tensor在gpu上，应先转化为cpu上后再进行转化

dtype = torch.cuda.FloatTensor
gpu_tensor = torch.randn(10, 20).type(dtype)

将tensor放到cpu上

cpu_tensor = gpu_tensor.cpu()

Tensor属性的访问

print(cpu_tensor.shape)    #形状
print(cpu_tensor.size())    #形状
print(cpu_tensor.type())    #数据类型
print(cpu_tensor.dim())   #维度
print(cpu_tensor.numel())    #元素个数

Tensor的操作

创造全一数组

x = torch.ones(2, 2)

创造随机数组

x = torch.randn(3,4)

查看/转换种类

x.type() #查看种类

x = x .long()    # 转化成整形
x = x.float()    #转回浮点型

算术操作

max_value, max_idx = torch.max(x, dim=0)  # 按列求最大值
sumx = torch.sum(x)    #求总和
sumy = torch.sum(x, dim=1) #按行求和

维度变化

x1 = x.unsqueeze(0)
x2 = x.unsqueeze(1)
x3 = x.squeeze()

改变形状

x = x.view(-1, 5) # -1 指随意取值，5是把第二维调成5
x = x.view(3, 20) # reshape (3, 20) 

Variable

from torch.autograd import Variable

Variable支持反向求导

x_tensor = torch.randn(10, 5)
y_tensor = torch.randn(10, 5)
# tensor转化成Variable
x = Variable(x_tensor, requires_grad=True) # 􁼕默认Variable不需要梯度 
y = Variable(y_tensor, requires_grad=True)

我们如果设定z为结果，使得z=x+2y

再将z的总和反向求导，注意，反向求导的对象应该是标量

z.backward()