From the model card
Fields below are copied from the tags and counters on the HuggingFace repository timm/vit_base_patch16_224.augreg_in21k at our last fetch. They are set by the uploader, not verified by us; rows with no tag are omitted. How this page is made.
- Publisher (HF namespace)
- timm
- Pipeline tag
- image-classification
- Library
- timm, Transformers
- Framework tags
- PyTorch
- Weight formats
- safetensors
- License tag
apache-2.0— read the license file in the repo before relying on it- Papers cited
- arXiv:2106.10270, arXiv:2010.11929
- Datasets declared
- imagenet-21k
- Downloads (HF counter at last fetch)
- 442,613
- Likes (HF counter at last fetch)
- 11
- Model card
- https://huggingface.co/timm/vit_base_patch16_224.augreg_in21k
Use cases
- Transfer learning backbone for custom image classification tasks
- Feature extraction for downstream retrieval or clustering pipelines
- Pre-trained initialisation for medical or satellite image fine-tuning
- Ablation studies comparing AugReg vs plain ImageNet-21k pre-training
- Teaching ViT fine-tuning workflows with a well-studied checkpoint
Pros
- AugReg pre-training yields better downstream transfer than standard ViT-B/16
- Apache 2.0 license; timm ecosystem provides comprehensive fine-tuning utilities
- Thoroughly benchmarked; performance numbers are easy to find in literature
- Patch16 is the most widely supported ViT configuration across frameworks
Cons
- ImageNet-21k labels contain noise; downstream tasks with different visual distributions may need more fine-tuning
- 224px resolution; high-resolution tasks require interpolating position embeddings
- ViT-Base is outperformed on most benchmarks by later architectures like ConvNeXt-B and DeiT-III
- timm dependency can conflict with other vision library requirements
Tags
timmpytorchsafetensorsimage-classificationtransformersdataset:imagenet-21karxiv:2106.10270arxiv:2010.11929license:apache-2.0region:us