DINOv2

DINOv2 frozen features can readily be used in models predicting per-pixel depth from a single image, both in and out-of-distribution.

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.

Visualization of the three first principal components of the patch features of all frames, mapped to RGB values.

Pretrained models

model# ofparamsImageNetk-NNImageNetlineardownloadViT-S/14 distilled21 M79.0%81.1%backbone onlyViT-B/14 distilled86 M82.1%84.5%backbone onlyViT-L/14 distilled300 M83.5%86.3%backbone onlyViT-g/141,100 M83.5%86.5%backbone only

Pretrained models via PyTorch Hub

Please follow the instructions here to install the PyTorch and torchvision dependencies (these are the only required dependencies). Installing both PyTorch and torchvision with CUDA support is strongly recommended.

The corresponding model card can be found in the [ MODEL_CARD.md] file.

``` 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')

```

Installation

The training and evaluation code requires PyTorch 2.0 and xFormers 0.0.18 as well as a number of other 3rd party packages. To setup all the required dependencies for training and evaluation, please follow the instructions below:

conda (Recommended) - Create and activate a dinov2 conda environment using the provided environment definition:

``` conda env create -f conda.yaml conda activate dinov2

```

pip - Use the provided requirements.txt to install the dependencies:

``` pip install -r requirements.txt

```

Data preparation

Expected contents for the ImageNet-1k data folder:

/test/ILSVRC2012_test_00000001.JPEG
/test/[..]
/test/ILSVRC2012_test_00100000.JPEG
/train/n01440764/n01440764_10026.JPEG
/train/[...]
/train/n15075141/n15075141_9993.JPEG
/val/n01440764/ILSVRC2012_val_00000293.JPEG
/val/[...]
/val/n15075141/ILSVRC2012_val_00049174.JPEG
/labels.txt

For ImageNet-22k, please adapt the Dataset object accordingly.

Training

Fast setup: training DINOv2 ViT-L/16 on ImageNet-1k

Run DINOv2 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 \ train.dataset_path=ImageNet:split=TRAIN:root=:extra=

```

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.

Long setup: training DINOv2 ViT-L/14 on ImageNet-22k

Run 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 \ train.dataset_path=ImageNet22k:root=:extra=

```

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.

Evaluation

The training code regularly saves the teacher weights. In order to evaluate the model, run the following evaluation on a single node:

k-NN classification on ImageNet-1k

``` python dinov2/run/eval/knn.py \ --config-file /config.yaml \ --pretrained-weights /eval/training_24999/teacher_checkpoint.pth \ --output-dir /eval/training_24999/knn \ --train-dataset ImageNet:split=TRAIN:root=:extra= \ --val-dataset ImageNet:split=VAL:root=:extra=

```

Logistic regression classification on ImageNet-1k

``` python dinov2/run/eval/log_regression.py \ --config-file /config.yaml \ --pretrained-weights /eval/training_24999/teacher_checkpoint.pth \ --output-dir /eval/training_24999/logreg \ --train-dataset ImageNet:split=TRAIN:root=:extra= \ --val-dataset ImageNet:split=VAL:root=:extra=

```

Linear classification with data augmentation on ImageNet-1k

``` python dinov2/run/eval/linear.py \ --config-file /config.yaml \ --pretrained-weights /eval/training_24999/teacher_checkpoint.pth \ --output-dir /eval/training_24999/linear \ --train-dataset ImageNet:split=TRAIN:root=:extra= \ --val-dataset ImageNet:split=VAL:root=:extra=

```

We release the weights from evaluating the different models:

modelImageNettop-1linear evaluationViT-S/14 distilled81.1%linear head weightsViT-B/14 distilled84.5%linear head weightsViT-L/14 distilled86.3%linear head weightsViT-g/1486.5%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=:extra= \ --val-dataset ImageNet:split=VAL:root=:extra=

```