How to download the backbone first and set the model path?

[lemonci]

Hello, since I am using this on a slurm cluster and once the job is submitted to slurm, it has not access to the internet. So I would like to know I can download the models for lpips in advance and set the model path of it? Thanks.

[Borda]

To use TorchMetrics’ LPIPS offline with a local model path, you can create the LPIPS similarity network by passing the model_path argument pointing to your downloaded weights file. Then, pass this local network instance to the PerceptualPathLength metric by the sim_net parameter.
Here’s a minimal code snippet to illustrate this:

from torchmetrics.image.lpips import _LPIPS, PerceptualPathLength

# Create LPIPS similarity network with local weights file
local_net = _LPIPS(net="alex", model_path="local/path/to/lpips_weights.pth")

# Use the local LPIPS network in PerceptualPathLength metric
ppl_metric = PerceptualPathLength(sim_net=local_net)

# Now you can use ppl_metric in your workflow without internet access
# ...

With this setup, you can freely specify any local path to your pre-downloaded model weights, making it easy to run LPIPS-based metrics offline on Slurm or other clusters.