PyTorch represents data as tensors—multi-dimensional arrays of a single data type—wrapped in a class that bundles operations and processing methods. This section covers setting up a working PyTorch environment using Anaconda and CUDA.

Anaconda Setup

Download Anaconda from https://www.anaconda.com/download/ and run the installer. Choose an installation path containing only English characters. Once installed, launch Anaconda Navigator from the Start menu (look for the green icon).

Environment management: virtualenv vs conda

• virtualenv creates isolated environments but inherits the system Python version; packages are typically installed via pip or conda.
• conda combines environment and package management: it can create environments with any Python version, handle dependencies across languages, and resolve package conflicts effectively.

The default evnironment base (root) confirms a successful installation. You can manage environments using the Anaconda Prompt.

Creating a Virtual Environment

Open any Anaconda command prompt.

List existing environments:

conda env list

Create a new environment with a specific Python version:

conda create -n torch_env python=3.11

Confirm with y. Avoid creating an empty environment (e.g., conda create -n torch_env) as it will lack a Python interpreter.

Configuring Mirror Channels (Optional)

Default channles point to overseas servers, which can be slow. Check current channels:

conda config --show
conda config --get

Available mirrors:

• Tsinghua: https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
• BFSU: https://mirrors.bfsu.edu.cn/anaconda/pkgs/main
• Alibaba: http://mirrors.aliyun.com/anaconda/pkgs/main

Add a channel:

conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main

It is safer to specify the channel during package installation rather than permanently modifying the configuration.

Installing PyTorch with CUDA

Activate the target environment:

conda activate torch_env

Check your NVIDIA driver and CUDA version:

nvidia-smi

Note the reported CUDA Version.

pip Method (Recommended)

pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

Replace cu118 with your CUDA version if different.

conda Method (Alternative)

Using a mirror for cudatoolkit:

conda install cudatoolkit=11.8 -c https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
conda install pytorch torchvision torchaudio -c https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/pytorch/win-64/

This approach has more potential dependency conflicts.

Verifying the Installation

Check installed packages:

conda activate torch_env
pip list | findstr torch

Launch Python and test GPU availability:

import torch
print(torch.cuda.is_available())

True indicates a successful GPU-enabled PyTorch installation.

PyCharm Integration

1. Configure Python Interpreter
   In PyCharm, go to File → Settings → Project → Python Interpreter. Add a new Conda environment or select an existing one. The environment path is typically C:\Users\<username>\.conda\envs\torch_env. The Conda executable is at <Anaconda_root>\Scripts\conda.exe.

2. Terminal Setup
   By default, the PyCharm terminal may activate the base environment. To change to your virtual environment:
   Locate the Anaconda Prompt shortcut, open its properties, and copy the target string starting from cmd.exe. For example:

   cmd.exe "/K" E:\Anaconda3\Scripts\activate.bat E:\Anaconda3
   

   Adjust the activate path to your environment:

   cmd.exe "/K" E:\Anaconda3\Scripts\activate.bat C:\Users\<username>\.conda\envs\torch_env
   

   In PyCharm: File → Settings → Tools → Terminal, paste the adjusted command into "Shell path".

   To prevent auto-activation of base globally:

   conda config --set auto_activate_base false
   

   Then you can manually enter an environment:

   conda activate torch_env
   

Running .ipynb Files in PyCharm

1. Install Jupyter in your environment

   conda activate torch_env
   conda install notebook
   

2. Set up an IPython kernel

   python -m ipykernel install --user --name torch_env_kernel --display-name "Python [conda env:torch_env]"
   

   List kernels:

   jupyter kernelspec list
   

   Remove unnecessary kernels:

   jupyter kernelspec remove python3
   

3. Set a Jupyter password (avoid token each time)

   jupyter notebook password
   

4. Launch the notebook server

   jupyter notebook
   

   The console prints a URL like http://localhost:8888/; you can use the token or the password set earlier.

5. Connect PyCharm Create or open a .ipynb file, click the Jupyter server configuration, and enter the server URL (e.g., http://localhost:8888/).

If a cell remains stuck on In[*], update tornado and ipykernel:

pip install tornado --upgrade
pip install ipykernel --upgrade

Troubleshooting: "Not a valid Win32 application"

An error like OSError: [WinError 193] %1 is not a valid Win32 application when importing torch often stems from DLL conflicts or corrupted installations. A clean environment often resolves it:

conda create -n fresh_torch python=3.9
conda activate fresh_torch
pip install numpy
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

Verify:

import torch
print(torch.cuda.is_available())

Removing a Virtual Environment

To delete an environment completely:

conda deactivate
conda env remove --name torch_env