LibreDDRNetm-sem / README.md
Xuban's picture
Initial upload
a103d39 verified
|
Raw History Blame Contribute Delete
2.8 kB
---
license: mit
library_name: libreyolo
pipeline_tag: image-segmentation
tags:
- image-segmentation
- semantic-segmentation
- ddrnet
- cityscapes
datasets:
- cityscapes
---
# LibreDDRNetm-sem
DDRNet-23 Cityscapes 19-class semantic segmentation weights, converted for LibreYOLO.
> **Cityscapes non-commercial notice.** These weights were trained on Cityscapes, whose license
> restricts the dataset and its derivatives to non-commercial use. See [Dataset terms](#dataset-terms).
## Source
Derived from [ydhongHIT/DDRNet](https://github.com/ydhongHIT/DDRNet) at commit de0db317c5af4b0946231f475cf74ca4f51b8ac2.
The source checkpoint is the official DDRNet-23 Cityscapes checkpoint linked from the upstream README
([Google Drive](https://drive.google.com/file/d/16viDZhbmuc3y7OSsUo2vhA7V6kYO0KX6), SHA-256 `66b68a4e61b2ab41fba5e967acd708ee3d9dd4771fb1e1a245a9a6b5bbd1db4e`), val mIoU 79.5 as reported upstream.
Copyright (c) 2020 xingkong. Licensed under the MIT License.
## Modifications
Converted with `weights/convert_ddrnet_weights.py` in the [LibreYOLO source repository](https://github.com/LibreYOLO/libreyolo).
Key rename only: the `model.` training-wrapper prefix is stripped and the `loss.criterion.weight` buffer is dropped.
Learned parameters are unchanged, including the auxiliary head used for training. `weights/parity_ddrnet.py`
checks the port against the official code with `max_abs_diff == 0`. The tensors are wrapped with LibreYOLO
checkpoint metadata (`model_family=ddrnet`, `size=m`, `task=semantic`, `nc=19`, evaluation canvas 1024x2048).
## Usage
```bash
pip install libreyolo
```
```python
from libreyolo import LibreYOLO
model = LibreYOLO("LibreDDRNetm-sem.pt")
result = model("street.jpg", save=True)
```
## License
MIT License, on an implied basis. The upstream repository links these weights from its README and states no separate terms for them, so LibreYOLO hosts them under the repository's MIT license on that implied basis. This is LibreYOLO's disclosed interpretation, not a license confirmed by the publisher. See the [LICENSE](./LICENSE) and [NOTICE](./NOTICE) files in this repository.
## Dataset terms
These weights were trained on [Cityscapes](https://www.cityscapes-dataset.com/). The [Cityscapes license](https://www.cityscapes-dataset.com/license/) permits distributing abstract derivative models from which the dataset cannot be recovered, and restricts the dataset and its derivatives to **non-commercial** use. Check those terms before using these pretrained weights commercially. The dataset is not redistributed in this repository. Weights trained from the ImageNet backbones (`LibreYOLO/LibreDDRNet{s,m,l}-imagenet`) or from scratch on your own data carry no Cityscapes term; a fine-tune started from this checkpoint inherits it.