LibreGTRx-sem

GTR-X Cityscapes semantic segmentation weights (19 classes) converted for LibreYOLO.

GTR support is being prepared for LibreYOLO v1.6.0. Earlier PyPI releases may not include this model family.

Usage

With a LibreYOLO version that includes GTR semantic segmentation:

from libreyolo import LibreYOLO

model = LibreYOLO("LibreGTRx-sem.pt")
result = model.predict("street.jpg")[0]
mask = result.semantic_mask.data          # (H, W) Cityscapes train IDs

Images are letterboxed into a 1024x2048 canvas and the network averages overlapping 1024px windows at a 768px stride, as upstream evaluates Cityscapes. A Cityscapes frame runs unresized in three windows.

The source release reports 83.6 mIoU on Cityscapes val with 1024px sliding windows (32.2M parameters). LibreYOLO has not re-measured that number; it is quoted from the source release.

Source

Official GTR implementation, source revision 782e737efe2e6437ac537fbdcee089673d3376c1.

Published checkpoint, weight repository revision 9fc62c8c2b2c976835d0f1c1ffc544dbc0f9e29f, SHA-256 ee4ade38e2e6e398110cbe909604566b221dbec87743d8f35e09ca2ca6093b55. Copyright (c) 2026 Intellindust-AI-Lab. The source code is MIT licensed and the publisher's weight repository explicitly declares MIT.

Modifications

Selected the EMA state dict and added LibreYOLO schema v1.0 metadata. Learned parameters and state-dict keys are unchanged. Training/optimizer state was removed. Conversion uses weights/convert_gtr_sem_weights.py in the LibreYOLO source repository.

Validation

Strict loading was checked, and on CPU the converted weights match the pinned upstream graph exactly (max abs diff 0.0) for a single 1024px window and for upstream's sliding-window inference on a 1024x2048 input, using the portable attention operators on both sides. Cityscapes mIoU, CUDA parity and GPU latency were not re-measured.

License

MIT. See LICENSE and NOTICE.