從今天起要從頭開始學習PyTorch了,在此記下筆記。
PyTorch 入門第一步
Tensor
import Torch as t
x = t.tensor(5,3)#生成5*3的矩陣
print(x.size())# 輸出x的維度
print(x.size()[0]) #輸出x的第0維
print(x.size(0))# 輸出x的第0維
y = t.rand(5,3)# 生成0-1之間的隨機數矩陣
# 加法
x + y
t.add(x,y)
result = t.rand(5,3)
t.add(x,y,out=result)
inplace
y.add(x)#不改變y的值
y.add_(x)# 改變y的值
函數名帶下劃線的會修改Tensor本身,x.add(y),x.t()會返回一個新的Tensor,而x不變,但x.add_(y)改變x。
Tensor 與 numpy
tensor轉numpy
a = t.tensor(5,3)
b = a.numpy()
numpy 轉 tensor
import numpy as np
a = np.ones(5)
b = t.from_numpy(a)
轉換之後,因爲tensor和numpy共享內存,所以一個改變另一個也會被改變。
AutoGrad自動微分
Variable
data:tensor
grad:存梯度,也是一個variable
gradfn:function,反向傳播計算輸入的梯度
from torch.autograd import Variable
x = Variable(t.ones(2,2),require_grad=True)
y = x.sum()
y.backward()
print(x.grad)
# 1, 1
# 1, 1
梯度在反向傳播過程中是累加的,故每次反向傳播都會累加之前的梯度,所以反向傳播之前要把梯度清0.
x.grad.data.zero_()
神經網絡
定義網絡
- 繼承nn.Module
- 實現nn.Module中forward方法
- 把具有可學習參數的層放在構造函數__init()__中(不可學習參數的層可放可不放)
import torch.nn as nn
import torch
import torch.nn.functional as F
from torch.autograd import Variable
class MyNet(nn.Module):
def __init__(self):
# nn.Module子類的函數必須在構造函數中執行父類的構造函數
super(MyNet, self).__init__()
self.first = nn.Conv2d(in_channels=3, out_channels=16, kernel_size=3)
self.second = nn.Conv2d(in_channels=16, out_channels=32, kernel_size=3)
self.Linear1 = nn.Linear(in_features=32*12*12, out_features=20)
self.Linear2 = nn.Linear(in_features=20, out_features=1)
def forward(self, x):
x = self.first(x)
x = F.relu(x)
x = self.second(x)
x = F.relu(x)
x = x.view(x.size()[0], -1)
x = self.Linear1(x)
x = self.Linear2(x)
return x
Net = MyNet()
print(Net)
定義了forward函數之後,backward會被自動實現。
# 網絡參數
params = list(Net.parameters())
print(len(params))
# Net.named_parameters()可返回可學習的參數和名稱
for name, parameters in Net.named_parameters():
print(name, ":", parameters.size())
forward函數的輸入輸出均是Variable
input = Variable(torch.rand((1, 3, 16, 16)))
out = Net(input)
print(out.size())
torch.nn只支持mini-natch,故輸入必須是4維的,
損失函數
target = torch.rand((1, 1))
criterion = nn.MSELoss()
loss = criterion(out, target)
print(loss)
Net.zero_grad()
print("反向傳播之前的梯度:", Net.first.bias.grad)
loss.backward()
print("反向傳播之後的梯度:", Net.first.bias.grad)
優化器
import torch.optim as optim
optimizer = optim.Adam(Net.parameters, lr=0.01)
# 梯度先清零
optimizer.zero_grad()
out = Net(input)
print(out.size())
target = torch.rand((1, 1))
criterion = nn.MSELoss()
loss = criterion(out, target)
loss.backward()
# 更新參數
optimizer.step()
數據加載與預處理
daraloader是可迭代的對象,將dataset返回的每一個樣本拼接成一個batch,並提供多行程加速優化和代碼打亂等。程序對dataset遍歷完一遍之後,對dataloader也完成了一次迭代。
import torchvision as tv
import torch as t
import torchvision.transforms as transforms
from torchvision.transforms import ToPILImage
import torch.nn as nn
import torch
import torch.nn.functional as F
from torch.autograd import Variable
transform = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5),(0.5, 0.5, 0.5)),])
trainset = tv.datasets.CIFAR10(
root='F:/code/EDSR-PyTorch-master-png/',
train=True, download=True, transform=transform)
trainloader = t.utils.data.DataLoader(
trainset,
batch_size=4,
shuffle=True,
num_workers=0
)
testset = tv.datasets.CIFAR10(
root='F:/code/EDSR-PyTorch-master-png/',
train=False, download=True, transform=transform)
testloader = t.utils.data.DataLoader(
testset,
batch_size=4,
shuffle=False,
num_workers=0
)