关于MacOS系统中PyTorch模型及torchvision来源PyTorch模型默认下载路径的技术问询

Let's break down your questions step by step—since you mentioned you didn't find any torch-related content in your home directory with ls -la ~/ | grep torch, that's totally expected! The default paths are tucked away in hidden directories, which your initial command wouldn't surface directly.

1. Default Download Path for PyTorch Core Models #

By default, PyTorch stores downloaded models (like those fetched via torch.hub.load()) in a hidden cache directory. If you haven't set the TORCH_HOME environment variable, the exact path is:
~/.cache/torch/hub/checkpoints/

You missed it earlier because .cache is a hidden folder—your ls -la ~/ command would show the .cache directory itself, not its contents. To check what's inside, run this command:

ls -la ~/.cache/torch/hub/checkpoints/

If you want to confirm the path programmatically, fire up a Python shell and run:

import torch
print(torch.hub.get_dir())

This will output the root hub directory, and your downloaded models live in the checkpoints subfolder under it.

2. Default Download Path for torchvision Models #

For models from torchvision (like ResNet, ViT, or EfficientNet), the default path aligns with PyTorch's hub cache if you're using newer versions (torchvision >= 0.13):
~/.cache/torch/hub/checkpoints/

Older torchvision versions might use a separate hidden directory:
~/.cache/torchvision/models/

To verify, you can either check the directory directly:

ls -la ~/.cache/torchvision/models/

Or run this Python snippet to confirm where your specific model gets saved:

from torchvision.models import resnet50
# Load a model with default weights (triggers download if not already cached)
model = resnet50(weights="DEFAULT")
# Print the root cache directory for torch models
print(torch.hub.get_dir())

Quick Tip: Override the Default Path #

If you want to move model downloads to a more visible location, set the TORCH_HOME environment variable in your shell config (like ~/.zshrc or ~/.bash_profile):

export TORCH_HOME="/path/to/your/custom/directory"

After adding this and restarting your shell, all PyTorch and torchvision models will download to /path/to/your/custom/directory/hub/checkpoints/ instead.

内容的提问来源于stack exchange，提问作者mrgloom