张量
引入pytorch,生成一个随机的5x3张量
>>> from __future__ import print_function
>>> import torch
>>> x = torch.rand(5, 3)
>>> print(x)
tensor([[0.5555, 0.7301, 0.5655],
[0.9998, 0.1754, 0.7808],
[0.5512, 0.8162, 0.6148],
[0.8618, 0.3293, 0.6236],
[0.2787, 0.0943, 0.2074]])
声明一个5x3的张量,张量中所有元素初始化为0
>>> x = torch.zeros(5, 3, dtype=torch.long)
从数据直接构造张量,这里的数据一般是python数组
>>> x = torch.tensor([5.5, 3])
>>> print(x)
tensor([5.5000, 3.0000])
从一个已有的tensor上类似创建新的张量,新、旧张量的形状和数据类型相同,除非对dtype进行了覆盖声明
>>> x = x.new_ones(5, 3, dtype=torch.double)
>>> print(x)
tensor([[1., 1., 1.],
[1., 1., 1.],
[1., 1., 1.],
[1., 1., 1.],
[1., 1., 1.]], dtype=torch.float64)
>>> y = torch.rand_like(x, dtype=torch.float)
>>> print(y)
tensor([[0.6934, 0.9637, 0.0594],
[0.0863, 0.6638, 0.4728],
[0.3416, 0.0892, 0.1761],
[0.6831, 0.6404, 0.8307],
[0.6254, 0.4180, 0.2174]])
张量的size,numpy里是shape
>>> print(x.size())
torch.Size([5, 3])
张量的操作
张量相加
>>> x=torch.rand(5, 3)
>>> y = torch.zeros(5, 3)
>>> print(x + y)
tensor([[0.8991, 0.9222, 0.2050],
[0.2478, 0.7688, 0.4156],
[0.4055, 0.9526, 0.2559],
[0.9481, 0.8576, 0.4816],
[0.0767, 0.3346, 0.0922]])
>>> print(torch.add(x, y))
tensor([[0.8991, 0.9222, 0.2050],
[0.2478, 0.7688, 0.4156],
[0.4055, 0.9526, 0.2559],
[0.9481, 0.8576, 0.4816],
[0.0767, 0.3346, 0.0922]])
>>> result = torch.empty(5, 3)
>>> torch.add(x, y, out=result)
tensor([[0.8991, 0.9222, 0.2050],
[0.2478, 0.7688, 0.4156],
[0.4055, 0.9526, 0.2559],
[0.9481, 0.8576, 0.4816],
[0.0767, 0.3346, 0.0922]])
>>> y.add_(x)
tensor([[0.8991, 0.9222, 0.2050],
[0.2478, 0.7688, 0.4156],
[0.4055, 0.9526, 0.2559],
[0.9481, 0.8576, 0.4816],
[0.0767, 0.3346, 0.0922]])
张量内元素访问形式和numpy保持一致,如输出张量y的第二维度上下标是1的所有元素
>>> print(y[:, 1])
tensor([0.9222, 0.7688, 0.9526, 0.8576, 0.3346])
iew函数改变tensor的形状,类似numpy的reshape
>>> x = torch.randn(4, 4)
>>> y = x.view(16) # 变成1x16的张量
>>> z = x.view(-1, 8) # 变成第二维度是8,第一维度自动计算的张量,结果是2x8的张量
>>> print(x.size(), y.size(), z.size())
torch.Size([4, 4]) torch.Size([16]) torch.Size([2, 8])
只有一个元素的向量,取这个元素
>>> x = torch.randn(1)
>>> print(x)
tensor([0.8542])
>>> print(x.item())
0.8541867136955261
转换成numpy数组
>>> x = torch.rand(5, 3)
>>> x.numpy()
array([[0.9320856 , 0.473859 , 0.6787642 ],
[0.14365482, 0.1112923 , 0.8280207 ],
[0.4609589 , 0.51031697, 0.15313298],
[0.18854082, 0.4548 , 0.49709243],
[0.8351501 , 0.6160053 , 0.61391556]], dtype=float32)
除CharTensor外,所有的cpu张量从numpy转换成tensor
import numpy as np
a = np.ones(5)
b = torch.from_numpy(a)
np.add(a, 1, out=a)
print(a)
print(b)
在cpu和gpu之间移动tensor,
if torch.cuda.is_available():
device = torch.device("cuda") # a CUDA device object
y = torch.ones_like(x, device=device) # 直接在GPU设备上创建
x = x.to(device) # or just use strings ``.to("cuda")``
z = x + y
print(z)
print(z.to("cpu", torch.double)) # ``.to`` can also change dtype together!
构建网络和损失函数
损失函数用来衡量输入和目标之间的距离
from __future__ import print_function
import torch
import torch.nn as nn
import torch.nn.functional as F
class Net(nn.Module):
## 定义了网络的结构
def __init__(self):
super(Net, self).__init__()
## input is channel 1, output 6 channels with 3x3 convulutionanl kernel
self.conv1 = nn.Conv2d(1, 6, 3)
self.conv2 = nn.Conv2d(6, 16, 3)
# an affine operation: y = Wx + b, # 6*6 from image dimension
self.fc1 = nn.Linear(16*6*6, 120)
self.fc2 = nn.Linear(120, 84)
self.fc3 = nn.Linear(84, 10)
## 前向传播,函数名必须是forward
def forward(self, x):
# Max pooling over a (2, 2) window
x = F.max_pool2d(F.relu(self.conv1(x)), (2, 2))
# If the size is a square you can only specify a single number
x = F.max_pool2d(F.relu(self.conv2(x)), 2)
x = x.view(-1, self.num_flat_features(x))
x = F.relu(self.fc1(x))
x = F.relu(self.fc2(x))
x = self.fc3(x)
return x
def num_flat_features(self, x):
size = x.size()[1:] # all dimensions except the batch dimension
num_features = 1
for s in size:
num_features *= s
return num_features
## 新建一个Net对象
net = Net()
print(net)
params = list(net.parameters())
print(len(params))
print(params[0].size()) # conv1's .weight
# 声明一个1x1x32x32的4维张量作为网络的输入
input = torch.randn(1, 1, 32, 32)
# input = torch.randn(1, 1, 32, 32)
output = net(input)
# net.zero_grad()
# out.backward(torch.randn(1, 10))
target = torch.randn(10)
target = target.view(1, -1)
criterion = nn.MSELoss()
loss = criterion(output, target)
print(loss)
print(loss.grad_fn) # MSELoss
print(loss.grad_fn.next_functions[0][0]) # Linear
print(loss.grad_fn.next_functions[0][0].next_functions[0][0]) # ReLU
网络的反向传播,为了反向传播损失(error)所做的只需要调用loss.backward()函数,如果没有清除已有的梯度,反向传播会累积梯度
调用loss.backward()函数,看以下conv1的bias的梯度在调用前后的差别。
net.zero_grad() # zeroes the gradient buffers of all parameters
print('conv1.bias.grad before backward')
print(net.conv1.bias.grad)
loss.backward()
print('conv1.bias.grad after backward')
print(net.conv1.bias.grad)
使用SGD更新权重
公式:weight = weight - learning_rate * gradient
可以用下面的torch代码实现
learning_rate = 0.01
for f in net.parameters():
f.data.sub_(f.grad.data * learning_rate)
但是torch已经实现了各种权重更新方式,比如SGD, Nesterov-SGD, Adam, RMSProp等,可以直接调用
import torch.optim as optim
# create your optimizer
optimizer = optim.SGD(net.parameters(), lr=0.01)
# in your training loop:
optimizer.zero_grad() # zero the gradient buffers
output = net(input)
loss = criterion(output, target)
loss.backward()
optimizer.step() # Does the update