Download torch+cu from scratch (painless version)
Article directory
- Download torch+cu from scratch (painless version)
-
- I. Introduction
- Second, the specific steps to configure the GPU version of torch
-
- 1, Check the Cuda version installed on the computer
- 2. Search the whl package name and version to be downloaded on the pytoch official website
- 3. Download the specified torch, torchvision, torchaudio three libraries
- 4. Install into the target deep learning environment
- 5. Test whether the installation is successful
I. Introduction
Since I got some new computers, I often need to configure a deep learning environment, and I feel cumbersome to search all the time, so I just record the search for myself, which is fast hh.
Second, the specific steps to configure the GPU version of torch
Generally speaking, the normal conda, direct download with pip will be sent directly, or it will be sent later, so you need to download whl directly and install it. It is not ruled out that there are children of destiny who have successfully passed pip+default-timeout.
1, Check the Cuda version installed on the computer
nvcc -V
You can see that I am CUDA11.6
If you don't have Cuda installed, you can refer to this blog:
Super detailed tutorial on deep learning environment configuration [Anaconda+PyTorch (GPU version)+CUDA+cuDNN]
If it is not an N card, it can only. . . I have used the money ability, and there is nothing I can do to help.
2. Search the whl package name and version to be downloaded on the pytoch official website
pytorch past historical version address: https://pytorch.org/get-started/previous-versions/
For example: I found that my cuda version is 11.6, so let's find out which torch versions are supported by cu11.6
Then I found it on the picture: I found that there is a line #CUDA 11.6 and a line of pip below it.
The content in pip is the package we want to download manually.
Mine is:
torch1.13.0+cu116 torch vision0.14.0+cu116 torchaudio==0.13.0
You can install torch, torchvision, torchaudio version suitable for your own cuda according to your own situation
What if the desired torch version does not match its own cuda? Then uninstall cuda and reinstall it ( cautiously ), this is a tutorial for reinstalling cuda: Uninstall and install CUDA under windows
3. Download the specified torch, torchvision, torchaudio three libraries
Address: https://download.pytorch.org/whl/cu113
Note: Remember to turn off Magic Internet.
Here I download torch’s whl package as a chestnut
. After clicking, ctrl+f input: mine is 11.6, so output cu116 and other versions Similarly,
in the second section, you can know that the torch version to be downloaded is: torch==1.13.0+cu116
and then I am python10, windows system, so I download this: torch-1.13.0+cu116-cp310-cp310-win_amd64 .whl
Note: cp310 is not 310 objects, it refers to the python version.
The same applies to the other three.
download as above
You can see that the one I downloaded is the same as the one on the pytorch official website. It cannot be said that it is irrelevant, but it is exactly the same.
4. Install into the target deep learning environment
Open Anconda Prompt and enter:
conda create -n pytorch13 python=3.10
conda activate pytorch13
pip install C:\Users\Admin\Downloads\torch-2.0.0+cu117-cp39-cp39-win_amd64.whl
pip install C:\Users\Admin\Downloads\torchvision-0.14.0+cu116-cp310-cp310-win_amd64.whl
pip install C:\Users\Admin\Downloads\torchaudio-0.13.0+cu117-cp39-cp39-win_amd64.whl
Note: The address after pip install becomes the absolute address of the whl you downloaded, and the order of installation is torch torchvision torchaudio
5. Test whether the installation is successful
python
import torch
print(torch.cuda.is_available())
Congratulations when printing True, you have successfully configured the torch+GPU deep learning environment!