---
license: bsd-3-clause
library_name: libreyolo
pipeline_tag: image-segmentation
datasets:
  - detection-datasets/coco
tags:
  - semantic-segmentation
  - deeplabv3
  - torchvision
  - pascal-voc
  - libreyolo
---

# LibreDeepLabv3r101-sem

DeepLabv3 semantic segmentation with dilated ResNet-101, output stride 8, repackaged for
LibreYOLO. It predicts background plus 20 Pascal VOC-named foreground classes
from a checkpoint trained on the matching COCO subset. This is DeepLabv3, not
DeepLabv3+; there is no decoder or CRF.

```python
from libreyolo import LibreYOLO

model = LibreYOLO("LibreDeepLabv3r101-sem.pt")
result = model.predict("image.jpg")
mask = result.semantic_mask.data
```

## Source

Derived from [pytorch/vision](https://github.com/pytorch/vision) at commit
[`336d36e8db990a905498c73933e35231876e28bc`](https://github.com/pytorch/vision/commit/336d36e8db990a905498c73933e35231876e28bc)
(torchvision v0.26.0). Copyright (c) Soumith Chintala 2016 and torchvision
contributors. The source implementation is BSD-3-Clause.

Official checkpoint: [deeplabv3_resnet101_coco-586e9e4e.pth](https://download.pytorch.org/models/deeplabv3_resnet101_coco-586e9e4e.pth)  
Bytes: `244545539`  
SHA-256: `586e9e4e203fcbf17e1ad45533d8d33ab133fc762bf03101c5dd743995c08c0d`  
Published mIoU / pixel accuracy: 67.4 / 92.4.

The published metrics use torchvision's aspect-preserving evaluation preset.
LibreYOLO uses a fixed 520x520 stretch deployment contract, followed by
ImageNet normalization and restoration of the output mask to the source
canvas, so end-to-end metrics can differ.

## Modifications and verification

Conversion removes only the training-time `aux_classifier.*` tensors and adds
LibreYOLO v1.0 checkpoint metadata. Every retained runtime tensor and state-dict
key is unchanged. The native 520x520 logits are bit-exact against the pinned
torchvision implementation before postprocessing (`max_abs_diff == 0.0`).

The fixed-shape deployment graph was also tested through LibreYOLO's unified
backend:

- ONNX Runtime CPU: 100% identical public mask pixels; maximum logit difference
  `1.34e-5`.
- TorchScript: bit-exact logits and 100% identical public mask pixels.
- OpenVINO CPU: 99.9994% identical public mask pixels using
  the runtime's default reduced-precision execution hint.
- TensorRT 10.16 FP32 on RTX 5070 Ti: 99.9986% identical public
  mask pixels.

The converted file has `235177707` bytes and SHA-256
`4575b7d5b1b70e9c67225ae76c00f552b29c2e54b07d55cfee8da218a9f41429`. See `docs/provenance/deeplabv3.md` and
`weights/convert_deeplabv3_weights.py` in the
[LibreYOLO source repository](https://github.com/LibreYOLO/libreyolo).

## License

The checkpoint publisher did not attach a separate per-object license file.
This mirror applies the releasing project's BSD-3-Clause license on an
**implied, not publisher-confirmed**, basis. Torchvision warns that pretrained
models may have licenses or terms derived from training data and that users
must determine whether they have permission for their use case. COCO
annotations are CC BY 4.0; source images retain their individual Flickr terms.
See [`LICENSE`](./LICENSE) and [`NOTICE`](./NOTICE).