ConvNeXt (base)

Intended use. ConvNeXt-Base is included as a representative modern supervised convolutional neural network (CNN). For insect classification, it can be used as a feature extractor with a classifier trained on insect labels, or fine-tuned on an insect dataset.

Architecture and training

This is a modern general purpose CNN model, it uses a hierarchical architecture with depthwise convoltions. The selected checkpoint was pretrained with supervision on ImageNet-22K and then fine-tuned on ImageNet-1K. This checkpoint has not been specifically fine-tuned for insect classification.

Inputs and outputs

Accepts RGB images, resized and centre-cropped to 224 × 224 pixels using the checkpoint’s preprocessing settings. These are included already.

The default classification head outputs scores for 1,000 imagenet classes. With the classification head removed, the model produces a 1,024-dimensional image embedding that can be used to train an insect classifier.

Performance

The authors report 85.8% top-1 accuracy on the ImageNet-1K validation set for ConvNeXt-Base pretrained on ImageNet-22K and fine-tuned and evaluated at 224 × 224 pixels.

Limitations

The supplied classifier is restricted to 1,000 ImageNet categories and cannot identify arbitrary insect taxa from their names. Classification using a different insect label set requires a trained classifier or fine-tuning with labeled examples.

How to obtain the weights

Publicly available on Hugging Face.

Citation

@inproceedings{liu2022convnet,
  title     = {A ConvNet for the 2020s},
  author    = {Liu, Zhuang and Mao, Hanzi and Wu, Chao-Yuan and
               Feichtenhofer, Christoph and Darrell, Trevor and Xie, Saining},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year      = {2022},
  url       = {https://arxiv.org/abs/2201.03545}
}

@misc{rw2019timm,
  title     = {PyTorch Image Models},
  author    = {Wightman, Ross},
  year      = {2019},
  publisher = {GitHub},
  doi       = {10.5281/zenodo.4414861},
  url       = {https://github.com/huggingface/pytorch-image-models}
}