nikhilr01/sen-bridge-weights

sen-bridge metric weights

Converted copies of the third-party weights used by the sen-bridge metrics (FID and LPIPS), so they can be loaded without PyTorch. Both the PyTorch and JAX backends load these exact files.

File	Contents	SHA-256
fid_inception.safetensors	FID Inception-v3 (pytorch-fid port of the TensorFlow inception-2015-12-05 weights)	b7a528bf4051f62c057bc14657865abf180a23eb1748d00ff0933d84698202ed
lpips_vgg.safetensors	torchvision VGG16 features trunk plus the LPIPS v0.1 linear layers	a96ef46bc0b0a064b67e277c15af82caa42278dd79aba0034815ed355e1aaacb

These are tensor-for-tensor conversions of the original files (the .pth pickles saved as safetensors, no retraining or quantisation). The original files' SHA-256 digests are pinned in sen_bridge/weights.py, which can also rebuild these files from the original hosts.

Licences (the weights are not the work of this repository's maintainer)

• FID Inception-v3: derived from the TensorFlow inception-2015-12-05 archive, Apache-2.0, Copyright 2015 The TensorFlow Authors, via pytorch-fid (Apache-2.0). See licenses/LICENSE-tensorflow-inception.txt and licenses/LICENSE-pytorch-fid.txt.
• LPIPS v0.1 linear layers: BSD-2-Clause, Copyright Richard Zhang et al. See licenses/LICENSE-lpips.txt.
• VGG16 trunk: from torchvision's vgg16-397923af ImageNet weights. torchvision publishes no separate licence for its weights and states that its pre-trained models may have their own licences or terms derived from the training data. No licence is asserted for this part. It is redistributed at the maintainer's decision and responsibility; if you are a rights holder and object, contact the maintainer and it will be removed. Check the terms yourself before commercial use. See licenses/LICENSE-torchvision.txt for torchvision's own code licence.

Maintainer: Nikhil Raghavendra.