swinv2_base_window12to24_192to384_ms_in22k_ft_in1k

See our collection for all versions of Swin Transformer V2.

Run Swin Transformer V2 with Keras 3: JAX, PyTorch, or TensorFlow

zeromodels/swinv2_base_window12to24_192to384_ms_in22k_ft_in1k

Paper: Swin Transformer V2: Scaling Up Capacity and Resolution (arXiv:2111.09883) · HF Papers

Swin V2 scales capacity and resolution with residual post-norm and scaled cosine attention. Classifier + hierarchical backbone.

For more details on the model, please go to the upstream model card.

Pure-Keras 3 conversion of timm/swinv2_base_window12to24_192to384.ms_in22k_ft_in1k for zeromodels. One implementation runs unmodified on TensorFlow / Torch / JAX.

This is an image-classification / backbone checkpoint (SwinV2ImageClassify / SwinV2Model).

✨ Quick start

import os

os.environ["KERAS_BACKEND"] = "torch"  # or "jax" / "tensorflow"

from PIL import Image
from zeromodels.models.swinv2 import SwinV2ImageClassify, SwinV2Model, SwinV2ImageProcessor

model = SwinV2ImageClassify.from_weights("zeromodels/swinv2_base_window12to24_192to384_ms_in22k_ft_in1k")
processor = SwinV2ImageProcessor.from_weights("zeromodels/swinv2_base_window12to24_192to384_ms_in22k_ft_in1k")

image = Image.open("your_image.jpg").convert("RGB")
pixels = processor(image)  # resize + normalize (normalization lives in the processor)
logits = model(pixels, training=False)
print(logits.shape)  # (1, num_classes)

# Feature extraction: the backbone without the classifier head
backbone = SwinV2Model.from_weights("zeromodels/swinv2_base_window12to24_192to384_ms_in22k_ft_in1k", as_backbone=True)
features = backbone(pixels, training=False)

Load any Swin Transformer V2 variant the same way with from_weights("zeromodels/<variant>"):

Tips

• Set KERAS_BACKEND before importing Keras / zeromodels.
• SwinV2ImageClassify returns class logits; SwinV2Model returns features (as_backbone=True for multi-scale stages).
• See docs and Loading Weights.
• Upstream / timm checkpoints: SwinV2ImageClassify.from_weights("hf:timm/swinv2_base_window12to24_192to384.ms_in22k_ft_in1k").

Special Thanks

A huge thank you to the Swin Transformer V2 authors and the timm / Hub communities for creating and releasing these models.

License: see YAML license (usually matches the upstream checkpoint).