适配CUDA10.1、Python3.7的Fastai/Torch/Torchvision版本及安装咨询

Hey there! Let's break down your questions and get you sorted out with the right versions and installation steps:

1. Compatible Versions for CUDA 10.1 & Python 3.7 #

After checking verified compatibility matrices and community feedback, here's the stable, tested combination for your setup:

• Fastai: 1.0.61 (the last major release of Fastai v1 that works seamlessly with the below PyTorch/torchvision versions)
• PyTorch: 1.3.1 (compiled specifically for CUDA 10.1)
• Torchvision: 0.4.2 (compiled specifically for CUDA 10.1)

This set ensures no version mismatches and full compatibility with your environment.

2. Step-by-Step Installation to Fix the CUDA Version Mismatch #

The error you're seeing is a classic case of PyTorch and torchvision being built against different CUDA versions. Let's fix this by doing a clean install of the matched versions:

1. Activate your new environment
   First, make sure you're working in the environment you created:

   • For Conda environments: conda activate your_env_name
   • For virtualenv (Linux/macOS): source your_env_name/bin/activate
   • For virtualenv (Windows): your_env_name\Scripts\activate

2. Uninstall conflicting packages
   Remove any existing versions of PyTorch, torchvision, and Fastai to avoid leftover files causing issues:

   pip uninstall -y torch torchvision fastai
   

   If you used Conda to install any of these, run conda remove -y torch torchvision fastai instead.

3. Install matched PyTorch & Torchvision (CUDA 10.1)
   Choose the command that fits your package manager:

   • For Conda: This command installs pre-compiled packages built explicitly for CUDA 10.1, ensuring perfect version alignment:

     conda install pytorch==1.3.1 torchvision==0.4.2 cudatoolkit=10.1 -c pytorch
     

   • For pip: The standardized wheel names will pull the correct CUDA 10.1-compatible versions for Python 3.7:

     pip install torch==1.3.1+cu101 torchvision==0.4.2+cu101
     

4. Install compatible Fastai
   Now install the Fastai version that pairs perfectly with these PyTorch/torchvision releases:

   pip install fastai==1.0.61
   

5. Verify your installation
   Run this quick check to confirm everything is working correctly:

   import torch
   import torchvision
   import fastai
   
   print(f"PyTorch version: {torch.__version__}")
   print(f"CUDA version in PyTorch: {torch.version.cuda}")
   print(f"Torchvision version: {torchvision.__version__}")
   print(f"Fastai version: {fastai.__version__}")
   

   You should see CUDA version 10.1 listed for PyTorch, and all version numbers matching what we installed. No more mismatch errors!

内容的提问来源于stack exchange，提问作者PremaK