Hiera Small 224 Hf

🚀 Hiera Model (Tiny, fine-tuned on IN1K)

Hiera is a hierarchical vision transformer that offers high speed, powerful performance, and above all, simplicity. It was introduced in the paper Hiera: A Hierarchical Vision Transformer without the Bells-and-Whistles and outperforms the state-of-the-art across a wide range of image and video tasks while being much faster.

📦 Dataset and Library Information

🚀 Quick Start

How does it work?

Vision transformers like ViT use the same spatial resolution and number of features throughout the whole network. However, this is inefficient as the early layers don't need that many features, and the later layers don't need that much spatial resolution. Prior hierarchical models like ResNet addressed this by using fewer features at the start and less spatial resolution at the end.

Several domain-specific vision transformers, such as Swin or MViT, have adopted this hierarchical design. But in the pursuit of state-of-the-art results using fully supervised training on ImageNet-1K, these models have become increasingly complicated as they add specialized modules to compensate for the spatial biases that ViTs lack. While these changes result in effective models with attractive FLOP counts, the added complexity actually makes these models slower overall.

We demonstrate that much of this complexity is actually unnecessary. Instead of manually adding spatial bases through architectural changes, we choose to teach the model these biases. By training with MAE, we can simplify or remove all of these bulky modules in existing transformers and increase accuracy in the process. The result is Hiera, an extremely efficient and simple architecture that outperforms the state-of-the-art in several image and video recognition tasks.

✨ Features

Intended uses & limitations

Hiera can be used for image classification, feature extraction, or masked image modeling. This specific checkpoint is intended for Feature Extraction.

💻 Usage Examples

Basic Usage

from transformers import AutoImageProcessor, HieraModel
import torch
from PIL import Image
import requests

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

image_processor = AutoImageProcessor.from_pretrained("facebook/hiera-small-224-hf")
model = HieraModel.from_pretrained("facebook/hiera-small-224-hf")

inputs = image_processor(images=image, return_tensors="pt")
outputs = model(**inputs)

Advanced Usage

You can also extract feature maps from different stages of the model using HieraBackbone and setting out_features when loading the model. This is how you would extract feature maps from every stage:

from transformers import AutoImageProcessor, HieraBackbone
import torch
from PIL import Image
import requests

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

image_processor = AutoImageProcessor.from_pretrained("facebook/hiera-small-224-hf")
# `out_features` should be a subset of ['stem', 'stage1', 'stage2', 'stage3', 'stage4']
# This introduce new LayerNorm layers and should probably train on a down-stream task
model = HieraBackbone.from_pretrained("facebook/hiera-small-224-hf", out_features=['stage1', 'stage2', 'stage3', 'stage4'])

inputs = image_processor(images=image, return_tensors="pt")
outputs = model(**inputs)
feature_maps = outputs.feature_maps

📚 Documentation

BibTeX entry and citation info

If you use Hiera or this code in your work, please cite:

@article{ryali2023hiera,
  title={Hiera: A Hierarchical Vision Transformer without the Bells-and-Whistles},
  author={Ryali, Chaitanya and Hu, Yuan-Ting and Bolya, Daniel and Wei, Chen and Fan, Haoqi and Huang, Po-Yao and Aggarwal, Vaibhav and Chowdhury, Arkabandhu and Poursaeed, Omid and Hoffman, Judy and Malik, Jitendra and Li, Yanghao and Feichtenhofer, Christoph},
  journal={ICML},
  year={2023}
}