Get familiar with pytoch and pytoch environment configuration

pytorch Development and framework of

        Deep learning framework , For me, it is divided into Tensorflow and Pytorch Two camps . The main difference between them is whether static graph takes precedence or dynamic graph takes precedence .
       

Dynamic graph static graph

        Dynamic graph ： Step by step , It has results and conforms to people's thinking .
        Static diagram ： Only the formula , Then finally input the data . This is a state of separation , Define the formula ,run When inputting data directly, it is used and cannot be modified , namely define→run The process of .
       

Performance comparison between the two

       Tensorflow1 and 2 Incompatibility made me laugh ,1 The advantages of 2 Unable to inherit .
       

pytorch What can be done ？

       pytorch Sure GPU Speed up , Automatic derivation , There are common network layers .
        For example, do a GPU The acceleration of ：
        First use here Cuda Why longer ？ Because some configurations need to be initialized , It takes time , The second time you can see that there is obviously less time .
        For derivation , The code is as follows ：

#  assignment , tell torch I need derivation 
x = torch.tensor(1.)
a = torch.tensor(1., requires_grad=True)
b = torch.tensor(2., requires_grad=True)
c = torch.tensor(3., requires_grad=True)
...
grads = autograd.grad(y, [a,b,c])
#  It's been worked out a,b,c Are the corresponding derivatives 

        Common network layer ：nn.Linear nn.Conv2d nn.LSTM And some activation functions “ function ”.
       

Environment configuration

        Because I just changed Y9000P, And is 32G Memory +512G+1T Solid state of , So I'm not afraid of anything ！（bushi）
        My configuration idea is Anaconda Build on .

1. install Anaconda（python It will also be installed incidentally , What needs to be noticed inside is path Adding options must be important ！！！）
2. Find the corresponding version of your graphics card , Because my driver has been installed , So I don't need to go to the official website to install or update , If necessary , The first Cuda Install well .
3. install Cudnn
4. stay Anaconda Of Prompt Create a file named Pytorch Environment
5. stay Pytorch Find the corresponding version on the official website , It's activated Pytorch After the environment , Run the corresponding statement （ Because after you choose the version on the official website , There will be corresponding running statements , Copy to the command line ）
6. I choose jupyter notebook To operate , Enter jupyter notebook You can jump to the corresponding interface .

       

Problems encountered

How to check your own graphics card version / How to see yourself CUDA edition ？
Solution ： open nvidia panel , Help includes .

WARNING conda.gateways.disk.delete:unlink_or_rename_to_trash(140)
Solution ：csdn

Solving environment: failed with initial frozen solve. Retrying with flexible solve.
Solution ： It may be that the corresponding limit time is too short ; Change to Tsinghua source ; Download the package directly , Offline local installation .（ Or you even Cudnn Not even loaded ）