From the model card
Fields below are copied from the tags and counters on the HuggingFace repository microsoft/swinv2-tiny-patch4-window16-256 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)
- microsoft
- Pipeline tag
- image-classification
- Library
- Transformers
- Framework tags
- PyTorch
- License tag
apache-2.0— read the license file in the repo before relying on it- Papers cited
- arXiv:2111.09883
- Datasets declared
- imagenet-1k
- Downloads (HF counter at last fetch)
- 406,259
- Likes (HF counter at last fetch)
- 14
- Model card
- https://huggingface.co/microsoft/swinv2-tiny-patch4-window16-256
Use cases
- Image classification with shifted-window attention at 256px input
- Fine-tuning base for object detection with Swin V2 backbone
- Ablation studies on window attention size effects
- Transfer learning to domain-specific image classification tasks
Pros
- Apache-2.0 license
- Swin V2 improvements over V1: better scale generalization via log-spaced position bias
- Transformers pipeline compatible
- Tiny size enables fast fine-tuning on consumer hardware
Cons
- 16-patch window at 256px is smaller than Swin V2's optimal configuration — some receptive field limitations
- Outperformed by ConvNeXt and ViT models at similar parameter counts on ImageNet
- Swin's shifted-window attention adds complexity to custom ONNX export
- Tiny variant significantly trails larger Swin V2 variants on dense prediction tasks
Tags
transformerspytorchswinv2image-classificationvisiondataset:imagenet-1karxiv:2111.09883license:apache-2.0endpoints_compatibleregion:usdeploy:azure