Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Patrick Labatut, Armand Joulin, Piotr Bojanowski
[Paper] [Blog] [Demo] [BibTeX]
PyTorch implementation and pretrained models for DINOv2. For details, see the paper: DINOv2: Learning Robust Visual Features without Supervision.
DINOv2 models produce high-performance visual features that can be directly employed with classifiers as simple as linear layers on a variety of computer vision tasks; these visual features are robust and perform well across domains without any requirement for fine-tuning. The models were pretrained on a dataset of 142 M images without using any labels or annotations.
video-reference+dinov2.mp4
| model | # of params |
ImageNet k-NN |
ImageNet linear |
download |
|---|---|---|---|---|
| ViT-S/14 distilled | 21 M | 79.0% | 81.1% | backbone only |
| ViT-B/14 distilled | 86 M | 82.1% | 84.5% | backbone only |
| ViT-L/14 distilled | 300 M | 83.5% | 86.3% | backbone only |
| ViT-g/14 | 1,100 M | 83.5% | 86.5% | backbone only |
Please follow the instructions here to install PyTorch (the only required dependency for loading the model). Installing PyTorch with CUDA support is strongly recommended.
A corresponding model card is included in the repository.
import torch
dinov2_vits14 = torch.hub.load('facebookresearch/dinov2', 'dinov2_vits14')
dinov2_vitb14 = torch.hub.load('facebookresearch/dinov2', 'dinov2_vitb14')
dinov2_vitl14 = torch.hub.load('facebookresearch/dinov2', 'dinov2_vitl14')
dinov2_vitg14 = torch.hub.load('facebookresearch/dinov2', 'dinov2_vitg14')The training and evaluation code requires PyTorch 2.0 and xFormers 0.0.18 as well as a number of other 3rd party packages. Note that the code has only been tested with the specified versions and also expects a Linux environment. To setup all the required dependencies for training and evaluation, please follow the instructions below:
conda (Recommended) - Clone the repository and then create and activate a dinov2 conda environment using the provided environment definition:
conda env create -f conda.yaml
conda activate dinov2pip - Clone the repository and then use the provided requirements.txt to install the dependencies:
pip install -r requirements.txtThe root directory of the dataset should hold the following contents:
<ROOT>/test/ILSVRC2012_test_00000001.JPEG<ROOT>/test/[..]<ROOT>/test/ILSVRC2012_test_00100000.JPEG<ROOT>/train/n01440764/n01440764_10026.JPEG<ROOT>/train/[...]<ROOT>/train/n15075141/n15075141_9993.JPEG<ROOT>/val/n01440764/ILSVRC2012_val_00000293.JPEG<ROOT>/val/[...]<ROOT>/val/n15075141/ILSVRC2012_val_00049174.JPEG<ROOT>/labels.txt
The provided dataset implementation expects a few additional metadata files to be present under the extra directory:
<EXTRA>/class-ids-TRAIN.npy<EXTRA>/class-ids-VAL.npy<EXTRA>/class-names-TRAIN.npy<EXTRA>/class-names-VAL.npy<EXTRA>/entries-TEST.npy<EXTRA>/entries-TRAIN.npy<EXTRA>/entries-VAL.npy
These metadata files can be generated (once) with the following lines of Python code:
from dinov2.data.datasets import ImageNet
for split in ImageNet.Split:
dataset = ImageNet(split=split, root="<ROOT>", extra="<EXTRA>")
dataset.dump_extra()Note that the root and extra directories do not have to be distinct directories.
Please adapt the dataset class to match your local setup.
dinov2 package should be included in the Python module search path, i.e. simply prefix the command to run with PYTHONPATH=..
Run DINOv2 training on 4 A100-80GB nodes (32 GPUs) in a SLURM cluster environment with submitit:
python dinov2/run/train/train.py \
--nodes 4 \
--config-file dinov2/configs/train/vitl16_short.yaml \
--output-dir <PATH/TO/OUTPUT/DIR> \
train.dataset_path=ImageNet:split=TRAIN:root=<PATH/TO/DATASET>:extra=<PATH/TO/DATASET>Training time is approximately 1 day and the resulting checkpoint should reach 81.6% on k-NN eval and 82.9% on linear eval.
The training code saves the weights of the teacher in the eval folder every 12500 iterations for evaluation.
Run DINOv2 training on 12 A100-80GB nodes (96 GPUs) in a SLURM cluster environment with submitit:
python dinov2/run/train/train.py \
--nodes 12 \
--config-file dinov2/configs/train/vitl14.yaml \
--output-dir <PATH/TO/OUTPUT/DIR> \
train.dataset_path=ImageNet22k:root=<PATH/TO/DATASET>:extra=<PATH/TO/DATASET>Training time is approximately 3.3 days and the resulting checkpoint should reach 82.0% on k-NN eval and 84.5% on linear eval.
The training code saves the weights of the teacher in the eval folder every 12500 iterations for evaluation.
The training code regularly saves the teacher weights. In order to evaluate the model, run the following evaluation on a single node:
python dinov2/run/eval/knn.py \
--config-file <PATH/TO/OUTPUT/DIR>/config.yaml \
--pretrained-weights <PATH/TO/OUTPUT/DIR>/eval/training_24999/teacher_checkpoint.pth \
--output-dir <PATH/TO/OUTPUT/DIR>/eval/training_24999/knn \
--train-dataset ImageNet:split=TRAIN:root=<PATH/TO/DATASET>:extra=<PATH/TO/DATASET> \
--val-dataset ImageNet:split=VAL:root=<PATH/TO/DATASET>:extra=<PATH/TO/DATASET>python dinov2/run/eval/log_regression.py \
--config-file <PATH/TO/OUTPUT/DIR>/config.yaml \
--pretrained-weights <PATH/TO/OUTPUT/DIR>/eval/training_24999/teacher_checkpoint.pth \
--output-dir <PATH/TO/OUTPUT/DIR>/eval/training_24999/logreg \
--train-dataset ImageNet:split=TRAIN:root=<PATH/TO/DATASET>:extra=<PATH/TO/DATASET> \
--val-dataset ImageNet:split=VAL:root=<PATH/TO/DATASET>:extra=<PATH/TO/DATASET>python dinov2/run/eval/linear.py \
--config-file <PATH/TO/OUTPUT/DIR>/config.yaml \
--pretrained-weights <PATH/TO/OUTPUT/DIR>/eval/training_24999/teacher_checkpoint.pth \
--output-dir <PATH/TO/OUTPUT/DIR>/eval/training_24999/linear \
--train-dataset ImageNet:split=TRAIN:root=<PATH/TO/DATASET>:extra=<PATH/TO/DATASET> \
--val-dataset ImageNet:split=VAL:root=<PATH/TO/DATASET>:extra=<PATH/TO/DATASET>We release the weights from evaluating the different models:
| model | ImageNet top-1 |
linear evaluation |
|---|---|---|
| ViT-S/14 distilled | 81.1% | linear head weights |
| ViT-B/14 distilled | 84.5% | linear head weights |
| ViT-L/14 distilled | 86.3% | linear head weights |
| ViT-g/14 | 86.5% | linear head weights |
The performance of the provided pretrained model weights can be evaluated as follows on ImageNet-1k:
python dinov2/run/eval/linear.py \
--config-file dinov2/configs/eval/vitg14_pretrain.yaml \
--pretrained-weights https://dl.fbaipublicfiles.com/dinov2/dinov2_vitg14/dinov2_vitg14_pretrain.pth \
--train-dataset ImageNet:split=TRAIN:root=<PATH/TO/DATASET>:extra=<PATH/TO/DATASET> \
--val-dataset ImageNet:split=VAL:root=<PATH/TO/DATASET>:extra=<PATH/TO/DATASET>DINOv2 code and model weights are released under the CC-BY-NC 4.0 license. See LICENSE for additional details.
See contributing and the code of conduct.
If you find this repository useful, please consider giving a star ⭐ and citation 🦖:
@misc{oquab2023dinov2,
title={DINOv2: Learning Robust Visual Features without Supervision},
author={Oquab, Maxime and Darcet, Timothée and Moutakanni, Theo and Vo, Huy V. and Szafraniec, Marc and Khalidov, Vasil and Fernandez, Pierre and Haziza, Daniel and Massa, Francisco and El-Nouby, Alaaeldin and Howes, Russell and Huang, Po-Yao and Xu, Hu and Sharma, Vasu and Li, Shang-Wen and Galuba, Wojciech and Rabbat, Mike and Assran, Mido and Ballas, Nicolas and Synnaeve, Gabriel and Misra, Ishan and Jegou, Herve and Mairal, Julien and Labatut, Patrick and Joulin, Armand and Bojanowski, Piotr},
journal={arXiv:2304.07193},
year={2023}
}