Pytorch environment configuration and basic knowledge

One 、 Configuration environment

Prepare in advance ： This article is installed by default Anaconda, The following operations are Anaconda Prompt on .
1, Creating a virtual environment ：conda create -n env_name python==version
2, Activate the virtual environment ：conda activate env_name
3, Look at the graphics driver ： nvidia-smi
4, To PyTorch Official website installation ： Suggest using CUDA edition
conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch
5, Verify successful installation ：
Input python Post input ：
import torch
torch.cuda.is_available()
CPU Version returns False,GPU Version returns True
6, Use PyCharm Connect to the virtual environment for use

Two 、 Basic knowledge of

1, tensor

tensor ： Generalization based on vector and matrix , From scalar , vector , matrix , Expand to higher dimensions .
for instance ：
0D： Scalar , Numbers
1D： vector
2D： matrix
3D： The time series , Single color picture
4D： Images
5D： video

stay PyTorch in ,torch.Tensor It is the main tool for storing and transforming data . And NumPy Quite similar . However ,Tensor Provide GPU Calculation and automatic gradient calculation, etc 更 multi-function , These make Tensor This data type is more suitable for deep learning .

2, Common operations

1, Randomly initialize the matrix ：torch.rand()
2, whole 0 matrix ：torch.zeros()
3, Build... Directly ：torch.tensor()
4, Addition operation ：torch.add(x,y)
5, Index operation ( Be similar to numpy): The indexed results share memory with the original data , To modify a , The other one will be modified . If you don't want to change it , Consider using copy() Other methods
6, Dimensional transformation ：
torch.view(): Back to the new tensor With the source tensor Shared memory ( It's actually the same tensor),更 Change one of them , The other one will change .( seeing the name of a thing one thinks of its function ,view() Just change 了 Yes, this one 量 Observation angle )
torch.reshape(): You can also change the shape of the tensor , However, this function does not guarantee to return its copy value , So the official 不 Recommended .
Recommended approach ： First use clone() Create a tensor copy and then use torch.view() Transform the function dimension . Use clone() Another advantage is that it will be recorded in the calculation diagram , That is, when the gradient is returned to the replica, it will also be transmitted to the source Tensor .

3, Broadcast mechanism

When two shapes 不 Identical Tensor When calculating by element , May trigger a broadcast (broadcasting) Mechanism ： First copy the elements appropriately so that the two Tensor After the shape is the same, operate according to the element .