LibreConvNeXtt-cls

ConvNeXt-Tiny image classifier (224px input, ImageNet-1k, 1000 classes), repackaged for LibreYOLO. ~28.6M parameters.

Source

Derived from huggingface/pytorch-image-models (timm), model convnext_tiny.fb_in1k. The ConvNeXt architecture is by Zhuang Liu et al. (Meta Platforms), originally released at facebookresearch/ConvNeXt under the MIT License. The ImageNet-1k pretrained weights are redistributed by timm under the Apache License 2.0 (no distillation, no extra data). Copyright (c) Meta Platforms, Inc. and the timm contributors.

ConvNeXt paper: A ConvNet for the 2020s (https://arxiv.org/abs/2201.03545).

Modifications

Learned parameters are unchanged from timm. The checkpoint is metadata-wrapped into the LibreYOLO format (model_family / task / nc / names). LibreYOLO's native ConvNeXt implementation mirrors timm's module names, so inference is bit-identical to timm (max_abs_diff == 0, verified by the parity test). See weights/convert_convnext_weights.py in the LibreYOLO source repository.

Usage

from libreyolo import LibreYOLO

# Auto-downloads from this repo on first use.
model = LibreYOLO("LibreConvNeXtt-cls.pt")
result = model.predict("image.jpg")[0]
print(result.probs.top1, result.probs.top5)   # ImageNet-1k class id + top-5 ids

License

Apache License 2.0. See the LICENSE and NOTICE files in this repository.

Only ConvNeXt V1 is published here. ConvNeXt-V2's small checkpoints are CC-BY-NC and are not redistributed.