LibreYOLO Classification
Collection
Image-classification model families packaged for LibreYOLO (ImageNet-1k, Apache-2.0). MobileNetV4-conv and ConvNeXt V1. • 26 items • Updated
LibreConvNeXtV2n-cls
NON-COMMERCIAL WEIGHTS: CC-BY-NC-4.0. The MIT architecture code has a separate license.
ConvNeXt V2 Nano, 224px, 1000 ImageNet classes, converted for LibreYOLO.
Usage
Requires LibreYOLO with ConvNeXt V2 support (the development branch until released).
from libreyolo import LibreYOLO
model = LibreYOLO("LibreConvNeXtV2n-cls.pt")
result = model.predict("image.jpg")[0]
print(result.names[result.probs.top1])
Source
https://github.com/facebookresearch/ConvNeXt-V2 at 2553895753323c6fe0b2bf390683f5ea358a42b9.
Copyright (c) Meta Platforms, Inc. and affiliates.
Official ImageNet-1K fine-tuned EMA checkpoint.
Source SHA-256: ac0c7985ee9e34f0b9e06f52abb6e72fa6127c42979191664901b15fcd185674.
Paper: ConvNeXt V2.
Modifications
LibreYOLO schema-v1 metadata and ImageNet class names added. Learned parameters
and state-dict keys are unchanged. Conversion: weights/convert_convnextv2_weights.py
in the LibreYOLO source.
Evaluation uses bicubic resize of the shorter side to 256, a 224 center crop,
and ImageNet mean/std normalization. This is the supervised classifier, not an
FCMAE encoder-only checkpoint. Fine-tunes derived from these weights retain
their non-commercial terms. No independent full-ImageNet accuracy claim is made.
License
Weights are CC-BY-NC-4.0. See LICENSE and NOTICE. The LICENSE reproduces the upstream combined MIT-code and CC-BY-NC-weight text.
- Downloads last month
- 70