Before you can start building deep learning models, you need to set up your development environment with PyTorch. This involves installing the core library and configuring your workspace. As this course assumes familiarity with Python, we'll focus on integrating PyTorch into a standard Python setup.

Choosing Your Installation Method: Conda vs. Pip

PyTorch can be installed using two primary package managers: Conda (part of the Anaconda or Miniconda distributions) and Pip (Python's default package installer).

Conda: Often preferred in data science and machine learning contexts. Conda manages packages and environments, simplifying the handling of complex dependencies, including non-Python libraries like the CUDA toolkit required for GPU acceleration. If you're already using Anaconda, this is typically the smoothest route.
Pip: The standard Python package installer. It works well, especially if you prefer managing your Python environments using tools like venv. You might need to handle the installation of NVIDIA drivers and the CUDA toolkit separately if you require GPU support.

For reproducibility and avoiding conflicts with other projects, it's highly recommended to install PyTorch within a dedicated virtual environment, regardless of whether you choose Conda or Pip.

Installation decision flowchart summarizing the choice between Conda and Pip, considering GPU requirements, and ending with verification.

Installation Steps

The exact installation command depends on your operating system (Linux, macOS, Windows), package manager (Conda, Pip), and whether you want CPU-only or GPU (CUDA-enabled) support.

The best practice is to always consult the official PyTorch website (pytorch.org) for the most up-to-date installation command generator. Select your preferences, and it will provide the precise command to run.

1. Using Conda

First, create and activate a new Conda environment (replace pytorch_env with your preferred name and choose a suitable Python version):

# Create the environment
conda create -n pytorch_env python=3.9 # Or 3.8, 3.10, etc.

# Activate the environment
conda activate pytorch_env

Now, go to the PyTorch website, select your OS, Conda, Python version, and desired CUDA version (or CPU). Copy the generated command. It will look something like this (example only, get the current one from the site!):

# Example Conda command for CUDA 11.8
conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia

# Example Conda command for CPU only
conda install pytorch torchvision torchaudio cpuonly -c pytorch

Run the command provided by the website in your activated Conda environment.

2. Using Pip

First, create and activate a new virtual environment using venv:

# Create the environment (in your project directory)
python -m venv pytorch_env

# Activate the environment
# Linux/macOS:
source pytorch_env/bin/activate
# Windows (Command Prompt):
.\pytorch_env\Scripts\activate
# Windows (PowerShell):
.\pytorch_env\Scripts\Activate.ps1

Next, go to the PyTorch website, select your OS, Pip, Python version, and desired CUDA version (or CPU). Copy the generated command. It will often involve specifying an index URL and potentially the CUDA version (example only, get the current one from the site!):

# Example Pip command for CUDA 11.8
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

# Example Pip command for CPU only
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu

Run the command provided by the website in your activated virtual environment. Ensure your pip is up-to-date (pip install --upgrade pip).

GPU Support (CUDA)

Deep learning computations, particularly matrix multiplications, are significantly faster on NVIDIA GPUs using the CUDA parallel computing platform.

Requirement: You need a CUDA-compatible NVIDIA GPU.
Drivers & Toolkit: You must have the appropriate NVIDIA display drivers and the CUDA Toolkit installed on your system before installing the CUDA-enabled version of PyTorch. The required CUDA Toolkit version must match the one PyTorch was built against (e.g., 11.7, 11.8, 12.1). Conda might handle the Toolkit installation for you, while with Pip, you usually need to install it system-wide first. Refer to NVIDIA's documentation and the PyTorch installation guide for specifics.
Installation Command: When using the PyTorch website's command generator, selecting a CUDA version will tailor the command to install the GPU-compatible binaries.

If you don't have a compatible GPU or don't need GPU acceleration initially, simply choose the "CPU" option during installation. You can perform all operations on the CPU, albeit potentially slower for large models.

Verifying the Installation

Once the installation completes, you can verify it by opening a Python interpreter (within your activated virtual environment) and running:

import torch

# Check PyTorch version
print(f"PyTorch Version: {torch.__version__}")

# Check if CUDA (GPU support) is available
cuda_available = torch.cuda.is_available()
print(f"CUDA Available: {cuda_available}")

if cuda_available:
    # Get the number of GPUs available
    print(f"Number of GPUs: {torch.cuda.device_count()}")
    # Get the name of the current GPU
    print(f"Current GPU Name: {torch.cuda.get_device_name(torch.cuda.current_device())}")
else:
    print("PyTorch is using CPU.")

# Create a simple tensor
x = torch.rand(2, 3)
print("Successfully created a tensor:")
print(x)

If these commands run without error and report the correct version and CUDA status (True if you installed the GPU version and have hardware/drivers set up, False otherwise), your installation is successful.

Environment Configuration Notes

Consistency: Always activate your dedicated pytorch_env (or whatever you named it) before working on your PyTorch projects or running scripts.
Common Libraries: You'll often use other libraries alongside PyTorch. Install them into the same environment using conda install <package> or pip install <package>:
- numpy: For numerical operations and interfacing with other libraries.
- matplotlib or seaborn: For plotting and visualization.
- scikit-learn: For traditional ML tasks and utilities.
- jupyterlab or notebook: For interactive development.
IDEs: Configure your Integrated Development Environment (IDE) like VS Code or PyCharm to use the Python interpreter located inside your virtual environment. This ensures code completion and debugging work correctly with PyTorch.

With PyTorch installed and your environment configured, you are now ready to explore its core component: the Tensor.