# Third-party notices

YellowCab uses third-party runtime software and pretrained initialization.
Those components remain subject to their respective licenses.

## Runtime dependencies

| Component | Version | License | Project |
|---|---:|---|---|
| NumPy | `2.2.6` | BSD 3-Clause | https://github.com/numpy/numpy |
| Pillow | `11.0.0` | HPND | https://github.com/python-pillow/Pillow |
| safetensors | `0.5.3` | Apache License 2.0 | https://github.com/huggingface/safetensors |
| PyTorch | `2.7.1` | BSD 3-Clause | https://github.com/pytorch/pytorch |
| torchvision | `0.22.1` | BSD 3-Clause | https://github.com/pytorch/vision |

## EfficientNet-B0 initialization

The image encoder was initialized from torchvision's
`EfficientNet_B0_Weights.IMAGENET1K_V1` weights. Torchvision documents these
weights as ported from Ross Wightman's EfficientNet implementation and trained
on ImageNet-1K:

- https://pytorch.org/vision/stable/models/generated/torchvision.models.efficientnet_b0.html
- https://github.com/rwightman/gen-efficientnet-pytorch

The repository distributes the resulting encoder parameters as part of the
YellowCab checkpoint. It does not distribute ImageNet images or taxi-training
images.

The torchvision BSD 3-Clause license is reproduced in
`THIRD_PARTY_LICENSES/torchvision-BSD-3-Clause.txt`. Ross Wightman's
EfficientNet implementation is licensed under Apache License 2.0; the license
text is reproduced in the top-level `LICENSE`.