pytorch实战3

1、tensor<--->numpy

import torch
import numpy as npx &＃61; np.array([[1,2],[3,4]])# 将numpy换成torch tensor
y &＃61; torch.from_numpy(x)# 将torch tensor换成numpy
z &＃61; y.numpy()

2、数据下载

import torch
import torchvision
from torchvision import transforms# 下载CIFAR-10数据集
train_dataset &＃61; torchvision.datasets.CIFAR10(root&＃61;&＃39;../../data/&＃39;,train&＃61;True,transform&＃61;transforms.ToTensor(),download&＃61;True
)# 取其中一组数据
image,label &＃61; train_dataset[0]
print(image.size())
print(label)train_loader &＃61; torch.utils.data.DataLoader(dataset &＃61; train_dataset,batch_size &＃61; 64,shuffle &＃61; True
)data_iter &＃61; iter(train_loader)images,labels &＃61; data_iter.next()for images,labels in train_loader:pass


3、梯度理解案例1

import torch
from torch.autograd import Variable# 创建tensor
x &＃61; torch.tensor(1.,requires_grad&＃61;True)
w &＃61; torch.tensor(2.,requires_grad&＃61;True)
b &＃61; torch.tensor(3.,requires_grad&＃61;True)# 创建一个计算图
y &＃61; w * x &＃43; b# 计算梯度
y.backward()# 输出梯度
print(w.grad) # --> x&＃61;1
print(x.grad) # --> w&＃61;2
print(b.grad) # --> 1&＃39;&＃39;&＃39;
计算谁的梯度就是将谁看作变量&＃xff0c;其他全部看成常量
&＃39;&＃39;&＃39;

4、梯度理解案例2

import torch
from torch.autograd import Variable# 输入
x &＃61; torch.randn(10,3)
# 输出
y &＃61; torch.randn(10,2)# 线性回归
linear &＃61; torch.nn.Linear(3,2)
print(linear.weight)
print(linear.bias)# 损失
loss_fn &＃61; torch.nn.MSELoss()# 学习率
learning_rate &＃61; 1e-3# 训练次数
epoch_n &＃61; 100# 优化器
optimiazer &＃61; torch.optim.Adam(linear.parameters(),learning_rate)for epoch in range(epoch_n):y_pred &＃61; linear(x)loss &＃61; loss_fn(y,y_pred)print("Eopch:{},Loss:{:.4f}".format(epoch,loss))linear.zero_grad()loss.backward()optimiazer.step()


5、预训练模型

import torchvision
import torch# 下载并加载预训练的ResNet-18
resnet &＃61; torchvision.models.resnet18(pretrained&＃61;True)# 如果你只想对模型的顶层进行微调&＃xff0c;请设置如下
for param in resnet.parameters():param.requires_grad &＃61; False# 替换顶层进行微调
resnet.fc &＃61; torch.nn.Linear(resnet.fc.in_features,100) # 100是个例子images &＃61; torch.randn(64,3,224,224)
outputs &＃61; resnet(images)print(outputs.size())# 保存以及载入模型
torch.save(resnet,&＃39;model.ckpt&＃39;)
model &＃61; torch.load(&＃39;model.ckpt&＃39;)# 仅仅保存和载入模型参数
torch.save(resnet.state_dict(),&＃39;params.ckpt&＃39;)
resnet.load_state_dict(torch.load(&＃39;params.ckpt&＃39;))