metadata

license: mit
tags:
  - vision
  - image-classification
datasets:
  - imagenet-21k
  - imagenet-1k
widget:
  - src: >-
      https://huggingface.co/datasets/mishig/sample_images/resolve/main/tiger.jpg
    example_title: Tiger
  - src: >-
      https://huggingface.co/datasets/mishig/sample_images/resolve/main/teapot.jpg
    example_title: Teapot
  - src: >-
      https://huggingface.co/datasets/mishig/sample_images/resolve/main/palace.jpg
    example_title: Palace

DiNAT (large variant with 11x11 kernel size)

DiNAT-Large with a 7x7 kernel pre-trained on ImageNet-21K at 224x224, and fine-tuned with 11x11 kernel size on ImageNet-1K at 384x384. It was introduced in the paper Dilated Neighborhood Attention Transformer by Hassani et al. and first released in this repository.

Model description

DiNAT is a hierarchical vision transformer based on Neighborhood Attention (NA) and its dilated variant (DiNA). Neighborhood Attention is a restricted self attention pattern in which each token's receptive field is limited to its nearest neighboring pixels. NA and DiNA are therefore sliding-window attention patterns, and as a result are highly flexible and maintain translational equivariance.

They come with PyTorch implementations through the NATTEN package.

Source

Intended uses & limitations

You can use the raw model for image classification. See the model hub to look for fine-tuned versions on a task that interests you.

Example

Here is how to use this model to classify an image from the COCO 2017 dataset into one of the 1,000 ImageNet classes:

from transformers import AutoImageProcessor, DinatForImageClassification
from PIL import Image
import requests

url = "http://images.cocodataset.org/val2017/000000039769.jpg"
image = Image.open(requests.get(url, stream=True).raw)

feature_extractor = AutoImageProcessor.from_pretrained("shi-labs/dinat-large-11x11-in22k-in1k-384")
model = DinatForImageClassification.from_pretrained("shi-labs/dinat-large-11x11-in22k-in1k-384")

inputs = feature_extractor(images=image, return_tensors="pt")
outputs = model(**inputs)
logits = outputs.logits
# model predicts one of the 1000 ImageNet classes
predicted_class_idx = logits.argmax(-1).item()
print("Predicted class:", model.config.id2label[predicted_class_idx])

For more examples, please refer to the documentation.

Requirements

Other than transformers, this model requires the NATTEN package.

If you're on Linux, you can refer to shi-labs.com/natten for instructions on installing with pre-compiled binaries (just select your torch build to get the correct wheel URL).

You can alternatively use pip install natten to compile on your device, which may take up to a few minutes. Mac users only have the latter option (no pre-compiled binaries).

Refer to NATTEN's GitHub for more information.

BibTeX entry and citation info

@article{hassani2022dilated,
    title        = {Dilated Neighborhood Attention Transformer},
    author       = {Ali Hassani and Humphrey Shi},
    year         = 2022,
    url          = {https://arxiv.org/abs/2209.15001},
    eprint       = {2209.15001},
    archiveprefix = {arXiv},
    primaryclass = {cs.CV}
}