LibreSwinl-cls

Swin Transformer V1 Large image classifier (224px input, ImageNet-1k, 1,000 classes), repackaged for LibreYOLO. 196,532,476 parameters.

Source

Derived from the Microsoft Research Swin Transformer V1 release at commit f82860bfb5225915aca09c3227159ee9e1df874d, with released weight storage at SwinTransformer/storage v1.0.0 (commit 3cc359915d3a6079b176a871f68d5fb0d8dfdea2). Copyright (c) 2021 SwinTransformer. Licensed under the MIT License.

The exact source snapshot is timm/swin_large_patch4_window7_224.ms_in22k_ft_in1k at revision e05e58ff5362edd212120dffaff3206a633c5534. Its model.safetensors SHA-256 is ccdcb5b425de65ed85875d5897681a72f7406c3fc07e087e933bace0c83807fd. Training lineage: ImageNet-22k pretraining followed by ImageNet-1k fine-tuning.

The Apache-2.0 timm v1.0.28 implementation at commit 8ef73809f622e0031bd7f4940265734aef8b9978 is the architecture and parity reference; it does not relicense the MIT weight release.

Modifications

Learned parameters are unchanged. The checkpoint is metadata-wrapped into the LibreYOLO schema with canonical ImageNet-1k class names. LibreYOLO's native Swin V1 implementation matches timm's parameter names and produces bit-exact pretrained logits (max_abs_diff == 0). The converted checkpoint SHA-256 is ad4d6efdd55b04c255482fa775240fd88a4097a8588be0be0633c6596ffbad88.

See weights/convert_swin_weights.py in the LibreYOLO source repository.

Usage

from libreyolo import LibreYOLO

model = LibreYOLO("LibreSwinl-cls.pt")
result = model.predict("image.jpg")[0]
print(result.probs.top1, result.probs.top5)

License

MIT License. See the LICENSE and NOTICE files in this repository.