Python——深度神经网络框架Pytorch安装与用法实例

深度神经网络框架 Pytorch
官网安装方法：
https://pytorch.org/get-started/locally/

测试代码 test.py

import torch
x = torch.rand(5, 3)
print("x =", x)
print("x.shape =",x.shape)
print("x.numpy() =", x.numpy())
print("x.tolist() =", x.tolist())

y = torch.rand(5, 3)
z = x + y
print("z =", z)

print("")
from torch.autograd import Variable
x_var=Variable(torch.ones(2,2),requires_grad=True)
print("x_var =", x_var)  
y_var=x_var.sum()
print("y_var =", y_var)
print("y_var.grad_fn =", y_var.grad_fn)
y_var.backward()
print("\n第一次反向传播, x_var.grad =",x_var.grad)
y_var.backward()
print("第二次反向传播, x_var.grad =",x_var.grad)
# 注意：grad在反向传播的过程中是累加的，深度学习是多层神经网络，在每次反向传播结束之后会累加上次的结果。
#  基于上述原因，训练过程中会在首次反向传播之前将梯度置为零。
 
#将梯度置为零
x_var.grad.data.zero_()
print(x_var.grad)
y_var.backward()
print("置零之后反向传播",x_var.grad)

运行 

python3 test.py 

输出为

x = tensor([[0.1214, 0.1568, 0.7282],
        [0.6820, 0.8997, 0.8014],
        [0.1159, 0.1559, 0.1364],
        [0.3021, 0.2610, 0.9011],
        [0.9494, 0.9833, 0.7934]])
x.shape = torch.Size([5, 3])
x.numpy() = [[0.12138915 0.15684658 0.7282294 ]
 [0.68202746 0.8996668  0.801434  ]
 [0.1159091  0.15591073 0.13638169]
 [0.30212373 0.26102382 0.9010761 ]
 [0.94941247 0.983263   0.79336965]]
x.tolist() = [[0.12138915061950684, 0.15684658288955688, 0.7282294034957886], 
[0.6820274591445923, 0.8996667861938477, 0.8014339804649353],
 [0.11590909957885742, 0.15591073036193848, 0.13638168573379517], 
[0.3021237254142761, 0.2610238194465637, 0.901076078414917], 
[0.94941246509552, 0.9832630157470703, 0.7933696508407593]]
z = tensor([[0.3496, 0.4020, 1.1696],
        [1.4339, 1.6175, 0.8582],
        [0.2218, 0.9734, 1.0614],
        [1.0951, 1.2197, 1.5408],
        [1.0147, 1.0386, 1.4496]])

x_var = tensor([[1., 1.],
        [1., 1.]], requires_grad=True)
y_var = tensor(4., grad_fn=<SumBackward0>)
y_var.grad_fn = <SumBackward0 object at 0x10bd18ca0>

第一次反向传播, x_var.grad = tensor([[1., 1.],
        [1., 1.]])
第二次反向传播, x_var.grad = tensor([[2., 2.],
        [2., 2.]])
tensor([[0., 0.],
        [0., 0.]])
置零之后反向传播 tensor([[1., 1.],
        [1., 1.]])