dvt-models / README.md
distant-viewing's picture
The terms as their sources state them, and where the TransNetV2 weights come from
1a21874 verified
|
Raw History Blame Contribute Delete
2.01 kB
---
license: other
license_name: mixed
tags:
- distant-viewing
- video
---
# Distant Viewing Toolkit models
The model weights that the [Distant Viewing Toolkit](https://github.com/distant-viewing/dvt)
loads and that are not published anywhere else. The toolkit fetches them from
here; the other models it uses come from their own repositories.
| File | Used by | What it is |
|-------------------------------------|------------|---------------------------------------------------------|
| `detect_shots.pt` | `frames` | TransNetV2 shot boundary weights, as a PyTorch state |
| `yolo26l_heads.pt` | `yolo` | a YOLO26-large detector trained to find heads |
| `laughter/laughter_net.pt` | `laughter` | the network of the laughter annotator |
| `laughter/laughter_classifier.json` | `laughter` | its classifier, 200 trees written out as json |
## Terms
The files do not share one license.
- `detect_shots.pt` holds the weights of TransNetV2 (Souček and Lokoč), which
its authors released under the MIT license at
[soCzech/TransNetV2](https://github.com/soCzech/TransNetV2). It is the
PyTorch state that the repository's `inference-pytorch/convert_weights.py`
makes from the weights under `inference/transnetv2-weights`. The copy is
kept here because that repository stores its weights in Git LFS, and its
quota is often used up.
- `yolo26l_heads.pt` was trained with the ultralytics library from its
`yolo26l.pt`, and is distributed under the AGPL-3.0 license as ultralytics
is. Ultralytics sells an Enterprise License for use without the AGPL's
conditions.
- The two laughter files were trained by the Distant Viewing Lab and are
released under the MIT license.
## Citation
Arnold, Taylor and Lauren Tilton. "Distant Viewing: Analyzing Large Visual
Corpora." *Digital Scholarship in the Humanities*, 2019.