pytorch中文文档:https://pytorch-cn.readthedocs.io/zh/latest/notes/autograd/
https://ptorch.com/docs/8/
一、动态图
https://blog.csdn.net/qjk19940101/article/details/79557866
二、变量
tensor的建立:
随机创立tensor:https://blog.csdn.net/dream161110/article/details/80293715 (rand,randn,normal,linespace的不同)
tensor的常规操作(包括建立和其他常用的):https://blog.csdn.net/xholes/article/details/81667211
三、网络结构
建立网络的四种方式:https://www.cnblogs.com/denny402/p/7593301.html
查看网络参数model.state_dict()
ps:
lenet模型例子。 https://blog.csdn.net/u010510350/article/details/77745183
argparse用来设置和解析参数。https://blog.csdn.net/lgczym/article/details/76474350
epsilon为算梯度时每次定义的梯度增加值(猜梯度?),太小会导致误差。 https://blog.csdn.net/freeliao/article/details/17564819
sequential建立网络时的特点:https://ptorch.com/news/57.html
四、网络流程
1.定义网络
2.定义结构
3.定义loss
4.设置优化器
例子:(来自莫烦python)
import torch import torch.nn.functional as F import matplotlib.pyplot as plt # torch.manual_seed(1) # reproducible x = torch.unsqueeze(torch.linspace(-1, 1, 100), dim=1) # x data (tensor), shape=(100, 1) y = x.pow(2) + 0.2*torch.rand(x.size()) # noisy y data (tensor), shape=(100, 1) # torch can only train on Variable, so convert them to Variable # The code below is deprecated in Pytorch 0.4. Now, autograd directly supports tensors # x, y = Variable(x), Variable(y) # plt.scatter(x.data.numpy(), y.data.numpy()) # plt.show() class Net(torch.nn.Module): def __init__(self, n_feature, n_hidden, n_output): super(Net, self).__init__() self.hidden = torch.nn.Linear(n_feature, n_hidden) # hidden layer self.predict = torch.nn.Linear(n_hidden, n_output) # output layer def forward(self, x): x = F.relu(self.hidden(x)) # activation function for hidden layer x = self.predict(x) # linear output return x net = Net(n_feature=1, n_hidden=10, n_output=1) # define the network print(net) # net architecture # SGD为随机梯度下降 optimizer = torch.optim.SGD(net.parameters(), lr=0.2) loss_func = torch.nn.MSELoss() # this is for regression mean squared loss plt.ion() # something about plotting for t in range(200): prediction = net(x) # input x and predict based on x loss = loss_func(prediction, y) # must be (1. nn output, 2. target) optimizer.zero_grad() # clear gradients for next train loss.backward() # backpropagation, compute gradients optimizer.step() # apply gradients #梯度在backward后就可以调用step,optimizer在之前需要先清零。step为单次优化 if t % 5 == 0: # plot and show learning process plt.cla() plt.scatter(x.data.numpy(), y.data.numpy()) plt.plot(x.data.numpy(), prediction.data.numpy(), 'r-', lw=5) plt.text(0.5, 0, 'Loss=%.4f' % loss.data.numpy(), fontdict={'size': 20, 'color': 'red'}) plt.pause(0.1) plt.ioff() plt.show()
with torch.no_grad(): 后面的代码不用反向传播
x.view(-1) 将tensor变成一维的(-1代表不确定)
五、细节问题
tensor相关函数的解析
https://www.jianshu.com/p/cf7adeff2a05
1.squeeze()和unsqueeze()函数,一个是增加维度,一个是减少维度
https://cloud.tencent.com/developer/article/1366479
2.detach()和detach_(),设置一个叶子节点,求导的终点
https://blog.csdn.net/u012436149/article/details/76714349
3. torch.linspace()
函数的作用是,返回一个一维的tensor(张量),这个张量包含了从start到end,分成steps个线段得到的向量。常用的几个变量
start:开始值
end:结束值
steps:分割的点数,默认是100
dtype:返回值(张量)的数据类型
https://blog.csdn.net/york1996/article/details/81671128
4. torch.sum()
将传入的整数代表的维度对应的数字相加减。
5. torch.argmax()
返回指定维度的最大元素的坐标
6. torch.clamp()
如超过最大值和最小值阀值,将其截断。numpy.clip作用类似
7. torch.cat()
合并多个tensor为一个tensor
8. tensor变量[range(number),...]
在range的维度中找后面list中一一对应的坐标(原来是tensor(n*m*p),就是range(n),list(length is m),),可以在对应维度空着。
tensors used as indices must be long or byte tensors. 如果是tensor需要改成long,或者直接用list就行,需要注意不能越界
ps:
torch.stack(tensors,dim) 将tensor维度叠加:https://blog.csdn.net/Teeyohuang/article/details/80362756
BN层以及BN问题:
https://blog.csdn.net/qq_25737169/article/details/79048516
https://www.cnblogs.com/guoyaohua/p/8724433.html
别人的pytorch总结:http://www.cnblogs.com/king-lps/p/8570021.html
强化学习:(很全,查td误差时候找到的)
https://blog.csdn.net/LagrangeSK/article/details/81239518
DRL各种DRL的pytorch代码:https://cloud.tencent.com/developer/article/1366479 (用Chrome打开)
pytorch调整学习率的六种方法:https://blog.csdn.net/shanglianlm/article/details/85143614