LibreSegformer weights
----------------------

These weights are derived from NVIDIA's SegFormer release
(https://github.com/NVlabs/SegFormer), specifically the checkpoint
published at https://huggingface.co/nvidia/segformer-b3-finetuned-ade-512-512.
Copyright (c) 2021, NVIDIA Corporation & affiliates. All rights reserved.

Licensed under the NVIDIA Source Code License for SegFormer. A complete copy
of that license is included in this repository as LICENSE. The license permits
redistribution of the work and of derivative works, provided the license
travels with them and attribution notices are retained, but it LIMITS USE TO
NON-COMMERCIAL PURPOSES ("research or evaluation purposes only", Section 3.3),
and Section 3.2 carries that limitation forward into every derivative work.

These weights are therefore NON-COMMERCIAL ONLY. They are NOT covered by
LibreYOLO's permissive license.

Modification: state-dict key remapping only (the upstream `segformer.` encoder
prefix becomes `encoder.`), plus LibreYOLO checkpoint metadata. Learned
parameters are NVIDIA's, unchanged; the converted checkpoint reproduces
upstream logits bit-exactly.

The model was fine-tuned on ADE20K (scene parsing, 150 classes), whose own
image terms restrict use to non-commercial research and education.

Citation:
  Xie, E., Wang, W., Yu, Z., Anandkumar, A., Alvarez, J. M., and Luo, P.
  "SegFormer: Simple and Efficient Design for Semantic Segmentation with
  Transformers." NeurIPS 2021.