You can deploy this model and start earning money today!

kerasformers/
swin_large_patch4_window12_384_ms_in22k_ft_in1k

Name: swin_large_patch4_window12_384_ms_in22k_ft_in1k
Availability: InStock

timm/swin_large_patch4_window12_384.ms_in22k_ft_in1k

finetune

mit

Views

No views yet

See our collection for all versions of Swin Transformer.

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

GitHub

zeromodels/swin_large_patch4_window12_384_ms_in22k_ft_in1k

Paper: Swin Transformer: Hierarchical Vision Transformer using Shifted Windows (arXiv:2103.14030) · HF Papers

Swin Transformer builds hierarchical feature maps with shifted-window attention. Strong as an ImageNet classifier and as a 4-stage backbone.

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

Pure-Keras 3 conversion of timm/swin_large_patch4_window12_384.ms_in22k_ft_in1k for zeromodels. One implementation runs unmodified on TensorFlow / Torch / JAX.

This is an image-classification / backbone checkpoint (SwinImageClassify / SwinModel).

✨ Quick start

python

import os

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

from PIL import Image
from zeromodels.models.swin import SwinImageClassify, SwinModel, SwinImageProcessor

model = SwinImageClassify.from_weights("zeromodels/swin_large_patch4_window12_384_ms_in22k_ft_in1k")
processor = SwinImageProcessor.from_weights("zeromodels/swin_large_patch4_window12_384_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 = SwinModel.from_weights("zeromodels/swin_large_patch4_window12_384_ms_in22k_ft_in1k", as_backbone=True)
features = backbone(pixels, training=False)

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

Variant	Hub
`swin_base_patch4_window12_384_ms_in1k`	`zeromodels/swin_base_patch4_window12_384_ms_in1k`
`swin_base_patch4_window12_384_ms_in22k`	`zeromodels/swin_base_patch4_window12_384_ms_in22k`
`swin_base_patch4_window12_384_ms_in22k_ft_in1k`	`zeromodels/swin_base_patch4_window12_384_ms_in22k_ft_in1k`
`swin_base_patch4_window7_224_ms_in1k`	`zeromodels/swin_base_patch4_window7_224_ms_in1k`
`swin_base_patch4_window7_224_ms_in22k`	`zeromodels/swin_base_patch4_window7_224_ms_in22k`
`swin_base_patch4_window7_224_ms_in22k_ft_in1k`	`zeromodels/swin_base_patch4_window7_224_ms_in22k_ft_in1k`
`swin_large_patch4_window12_384_ms_in22k`	`zeromodels/swin_large_patch4_window12_384_ms_in22k`
`swin_large_patch4_window12_384_ms_in22k_ft_in1k`	`zeromodels/swin_large_patch4_window12_384_ms_in22k_ft_in1k`
`swin_large_patch4_window7_224_ms_in22k`	`zeromodels/swin_large_patch4_window7_224_ms_in22k`
`swin_large_patch4_window7_224_ms_in22k_ft_in1k`	`zeromodels/swin_large_patch4_window7_224_ms_in22k_ft_in1k`
`swin_small_patch4_window7_224_ms_in1k`	`zeromodels/swin_small_patch4_window7_224_ms_in1k`
`swin_small_patch4_window7_224_ms_in22k`	`zeromodels/swin_small_patch4_window7_224_ms_in22k`
`swin_small_patch4_window7_224_ms_in22k_ft_in1k`	`zeromodels/swin_small_patch4_window7_224_ms_in22k_ft_in1k`
`swin_tiny_patch4_window7_224_ms_in1k`	`zeromodels/swin_tiny_patch4_window7_224_ms_in1k`
`swin_tiny_patch4_window7_224_ms_in22k`	`zeromodels/swin_tiny_patch4_window7_224_ms_in22k`

Tips

Set KERAS_BACKEND before importing Keras / zeromodels.
SwinImageClassify returns class logits; SwinModel returns features (as_backbone=True for multi-scale stages).
See docs and Loading Weights.
Upstream / timm checkpoints: SwinImageClassify.from_weights("hf:timm/swin_large_patch4_window12_384.ms_in22k_ft_in1k").

Special Thanks

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

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

kerasformers/swin_large_patch4_window12_384_ms_in22k_ft_in1k