This model performs visual feature extraction. For instance, RADIO generates image embeddings that can be used by a downstream model to classify images.

This model is for research and development only.

NVIDIA Research

References

[Paper] [PHI-S Paper] [BibTex][GitHub examples]

Model Architecture:

Architecture Type: Neural Network
Network Architecture: Vision Transformer

Input:

Input Type(s): Image
Input Format(s): Red, Green, Blue (RGB)
Input Parameters: Two Dimensional (2D)
Other Properties Related to Input: Image resolutions up to 2048x2028 in increments of 16 pixels

Output:

Output Type(s): Embeddings
Output Format: Tensor
Output Parameters: 2D
Other Properties Related to Output: Downstream model required to leverage image features

Software Integration:

Runtime Engine(s):

TAO- 24.10

Supported Hardware Microarchitecture Compatibility:

NVIDIA Ampere
NVIDIA Blackwell
NVIDIA Jetson
NVIDIA Hopper
NVIDIA Lovelace
NVIDIA Pascal
NVIDIA Turing
NVIDIA Volta

[Preferred/Supported] Operating System(s):

Linux
Linux 4 Tegra
QNX
Windows

License/Terms of Use

RADIO code and weights are released under the NSCLv1 License.

Pretrained Models

Refer to model_results.csv for model versions and their metrics.

Link: https://huggingface.co/collections/nvidia/radio-669f77f1dd6b153f007dd1c6

HuggingFace Hub

In order to pull the model from HuggingFace, you need to be logged in:

huggingface-cli login

Then you can pull the model from a Python script:

Python

from transformers import AutoModel
model = AutoModel.from_pretrained("nvidia/RADIO", trust_remote_code=True)

Alternatively, you can specify an access token:

Python

access_token = "<YOUR ACCESS TOKEN"
model = AutoModel.from_pretrained("nvidia/RADIO", trust_remote_code=True, token=access_token)

Usage

RADIO will return a tuple with two tensors. The summary is similar to the cls_token in ViT and is meant to represent the general concept of the entire image. It has shape $(B,C)$ with $B$ being the batch dimension, and $C$ being some number of channels. The spatial_features represent more localized content which should be suitable for dense tasks such as semantic segmentation, or for integration into an LLM. It has shape $(B,T,D)$ with $T$ being the flattened spatial tokens, and $D$ being the channels for spatial features. Note that $C \neq D$ in general.

Converting to a spatial tensor format can be done using the downsampling size of the model, combined with the input tensor shape. For 'radio_v1', the patch size is 14.

Python

from einops import rearrange
spatial_features = rearrange(spatial_features, 'b (h w) d -> b d h w', h=x.shape[-2] // patch_size, w=x.shape[-1] // patch_size)

The resulting tensor will have shape $(B,D,H,W)$, as is typically seen with computer vision models.

RADIOv1 Notes

We have trained this model to be flexible in input dimension. It supports inputs with both width and height in the range $[14, 1008]$ as long as both axes are divisible by 14. We have found that summarization tokens work best at $H=W=378$ (although the range $[192, 448]$ works well). For spatial tasks, we used $H=W=518$ to perform linear probing for semantic segmentation, and may perform better for more high-resolution tasks. Going up to $1008$, the model may need additional fine tuning at that resolution for best results.

It is not required that $H=W$ although we have not specifically trained or testing the model in this setting.

Training, Testing, and Evaluation Datasets:

Training Dataset:

Link: https://www.datacomp.ai/
** Data Collection Method by dataset

Automated
** Labeling Method by dataset
Not Applicable (no labels are needed)
Properties (Quantity, Dataset Descriptions, Sensor(s)): 12.8 billion diverse images gathered from the Internet using Common Crawl

Evaluation Dataset:

Link: ImageNet
** Data Collection Method by dataset

Automated
** Labeling Method by dataset
Human

Properties (Quantity, Dataset Descriptions, Sensor(s)): This dataset spans 1000 object classes and contains 1,281,167 training images, 50,000 validation images and 100,000 test images.

Inference:

Engine: PyTorch
Test Hardware: A100

Citing RADIO

If you find this repository useful, please consider giving a star and citation:

@InProceedings{Ranzinger_2024_CVPR,
    author    = {Ranzinger, Mike and Heinrich, Greg and Kautz, Jan and Molchanov, Pavlo},
    title     = {AM-RADIO: Agglomerative Vision Foundation Model Reduce All Domains Into One},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2024},
    pages     = {12490-12500}
}

@misc{ranzinger2024phisdistributionbalancinglabelfree,
      title={PHI-S: Distribution Balancing for Label-Free Multi-Teacher Distillation}, 
      author={Mike Ranzinger and Jon Barker and Greg Heinrich and Pavlo Molchanov and Bryan Catanzaro and Andrew Tao},
      year={2024},
      eprint={2410.01680},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2410.01680}, 
}

Ethical Considerations (For NVIDIA Models Only):

NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. When downloaded or used in accordance with our terms of service, developers should work with their internal model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse.