Hiera Abswin Base Mim

🚀 hiera_abswin_base_mim Model Card

This is an image encoder named Hiera, which uses an absolute window position embedding strategy and is pre - trained via Masked Image Modeling (MIM). This model has not been fine - tuned for a specific classification task. It is designed to serve as a general - purpose feature extractor or a backbone for downstream tasks such as object detection, segmentation, or custom classification.

🚀 Quick Start

The model can be used for image embeddings and detection feature map extraction. See the "💻 Usage Examples" section for detailed code examples.

✨ Features

• General - Purpose Feature Extractor: Can be used as a general - purpose tool for extracting image features.
• Backbone for Downstream Tasks: Suitable as a backbone for various downstream tasks like object detection, segmentation, or custom classification.

📦 Installation

The original README does not provide installation steps, so this section is skipped.

💻 Usage Examples

Basic Usage

Image Embeddings

import birder
from birder.inference.classification import infer_image

(net, model_info) = birder.load_pretrained_model("hiera_abswin_base_mim", inference=True)

# Get the image size the model was trained on
size = birder.get_size_from_signature(model_info.signature)

# Create an inference transform
transform = birder.classification_transform(size, model_info.rgb_stats)

image = "path/to/image.jpeg"  # or a PIL image
(out, embedding) = infer_image(net, image, transform, return_embedding=True)
# embedding is a NumPy array with shape of (1, 768)

Advanced Usage

Detection Feature Map

from PIL import Image
import birder

(net, model_info) = birder.load_pretrained_model("hiera_abswin_base_mim", inference=True)

# Get the image size the model was trained on
size = birder.get_size_from_signature(model_info.signature)

# Create an inference transform
transform = birder.classification_transform(size, model_info.rgb_stats)

image = Image.open("path/to/image.jpeg")
features = net.detection_features(transform(image).unsqueeze(0))
# features is a dict (stage name -> torch.Tensor)
print([(k, v.size()) for k, v in features.items()])
# Output example:
# [('stage1', torch.Size([1, 96, 56, 56])),
#  ('stage2', torch.Size([1, 192, 28, 28])),
#  ('stage3', torch.Size([1, 384, 14, 14])),
#  ('stage4', torch.Size([1, 768, 7, 7]))]

📚 Documentation

Model Details

🔧 Technical Details

The model uses a Hiera architecture with an absolute window position embedding strategy and is pre - trained using Masked Image Modeling (MIM). It has not been fine - tuned for specific classification tasks and is suitable for general - purpose feature extraction and as a backbone for downstream tasks.

📄 License

This project is licensed under the Apache 2.0 license.

📚 Citation

@misc{ryali2023hierahierarchicalvisiontransformer,
      title={Hiera: A Hierarchical Vision Transformer without the Bells-and-Whistles},
      author={Chaitanya Ryali and Yuan-Ting Hu and Daniel Bolya and Chen Wei and Haoqi Fan and Po-Yao Huang and Vaibhav Aggarwal and Arkabandhu Chowdhury and Omid Poursaeed and Judy Hoffman and Jitendra Malik and Yanghao Li and Christoph Feichtenhofer},
      year={2023},
      eprint={2306.00989},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2306.00989},
}

@misc{bolya2023windowattentionbuggedinterpolate,
      title={Window Attention is Bugged: How not to Interpolate Position Embeddings},
      author={Daniel Bolya and Chaitanya Ryali and Judy Hoffman and Christoph Feichtenhofer},
      year={2023},
      eprint={2311.05613},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2311.05613},
}