LibreDeiTs-cls

Plain DeiT-Small patch-16 image classifier (fixed 224px input, ImageNet-1k, 1000 classes), packaged for LibreYOLO. It has 22.1M parameters.

Source

The unchanged learned tensors come from timm/deit_small_patch16_224.fb_in1k at revision 91327a9c99f98fe6b524cd4d397b7226b80e1365. The source model.safetensors SHA-256 is 1e747b4a8d0df2cfbd3c450e8c97685d867448ab0c2ddbfb34b6885f5cb23e5b. That repository declares the checkpoint under Apache License 2.0.

The native graph follows huggingface/pytorch-image-models commit e98c05a5a15e81188ec62dd5380b8f5c3251075a (Apache-2.0). The original DeiT architecture is from facebookresearch/deit commit 7e160fe43f0252d17191b71cbb5826254114ea5b (Apache-2.0).

Modifications

Learned parameters and state-dict keys are unchanged. Conversion only adds LibreYOLO v1.0 checkpoint metadata and canonical ImageNet-1k class names. The native graph strict-loads the source state dictionary and produces bit-identical eager logits (max_abs_diff == 0). See weights/convert_deit_weights.py and docs/provenance/deit.md in the LibreYOLO source repository.

This museum model is inference-only. Distillation-token and 384px variants are not included.

Usage

from libreyolo import LibreYOLO

model = LibreYOLO("LibreDeiTs-cls.pt")  # downloads this checkpoint once
result = model.predict("image.jpg")
print(result.names[result.probs.top1], result.probs.top5)

License

Apache License 2.0. See LICENSE and NOTICE.