DINOV2

DINOv2: Learning Robust Visual Features without Supervision

This is a warehouse for DinoV2-Pytorch-model, can be used to train your image-datasets for classification tasks.
The code for the self-supervised training part is being improved and will be uploaded soon.
The code mainly comes from official source code.
Added DINOv2 backbones with registers, following Vision Transformers Need Registers.

Preparation

Download vit models pretrained_weights

pretrained_weights_website

Download the dataset:

flower_dataset.

Project Structure

├── datasets: Load datasets
    ├── masking.py: Construct data masking for self-supervise learning
    ├── my_dataset.py: Customize reading data sets and define transforms data enhancement methods
    ├── split_data.py: Define the function to read the image dataset and divide the training-set and test-set
    ├── threeaugment.py: Additional data augmentation methods
├── models: DinoV2 Model
    ├── build_models.py: Combine DinoV2 and dino_head as a cls model for image-classification
    ├── dino_head.py: Construct cls_head with DinoV2 model
    ├── dinov2.py: Construct DinoViT model
├── util:
    ├── engine.py: Function code for a training/validation process
    ├── losses.py: Knowledge distillation loss, combined with teacher model (if any)
    ├── optimizer.py: Define Sophia optimizer
    ├── samplers.py: Define the parameter of "sampler" in DataLoader
    ├── utils.py: Record various indicator information and output and distributed environment
├── estimate_model.py: Visualized evaluation indicators ROC curve, confusion matrix, classification report, etc.
└── train_gpu.py: Training model startup file

Precautions

Before you use the code to train your own data set, please first enter the train_gpu.py file and modify the data_root, batch_size, num_workers and nb_classes parameters. If you want to draw the confusion matrix and ROC curve, you only need to set the predict parameter to True

Use Sophia Optimizer (in util/optimizer.py)

You can use anther optimizer sophia, just need to change the optimizer in train_gpu.py, for this training sample, can achieve better results

# optimizer = create_optimizer(args, model_without_ddp)
optimizer = SophiaG(model.parameters(), lr=2e-4, betas=(0.965, 0.99), rho=0.01, weight_decay=args.weight_decay)

Train this model

Parameters Meaning:

1. nproc_per_node: <The number of GPUs you want to use on each node (machine/server)>
2. CUDA_VISIBLE_DEVICES: <Specify the index of the GPU corresponding to a single node (machine/server) (starting from 0)>
3. nnodes: <number of nodes (machine/server)>
4. node_rank: <node (machine/server) serial number>
5. master_addr: <master node (machine/server) IP address>
6. master_port: <master node (machine/server) port number>

Note:

If you want to use multiple GPU for training, whether it is a single machine with multiple GPUs or multiple machines with multiple GPUs, each GPU will divide the batch_size equally. For example, batch_size=4 in my train_gpu.py. If I want to use 2 GPUs for training, it means that the batch_size on each GPU is 4. Do not let batch_size=1 on each GPU, otherwise BN layer maybe report an error. If you recive an error like "error: unrecognized arguments: --local-rank=1" when you use distributed multi-GPUs training, just replace the command "torch.distributed.launch" to "torch.distributed.run".

train model with single-machine single-GPU：

python train_gpu.py

train model with single-machine multi-GPU：

python -m torch.distributed.launch --nproc_per_node=8 train_gpu.py

train model with single-machine multi-GPU:

(using a specified part of the GPUs: for example, I want to use the second and fourth GPUs)

CUDA_VISIBLE_DEVICES=1,3 python -m torch.distributed.launch --nproc_per_node=2 train_gpu.py

train model with multi-machine multi-GPU:

(For the specific number of GPUs on each machine, modify the value of --nproc_per_node. If you want to specify a certain GPU, just add CUDA_VISIBLE_DEVICES= to specify the index number of the GPU before each command. The principle is the same as single-machine multi-GPU training)

On the first machine: python -m torch.distributed.launch --nproc_per_node=1 --nnodes=2 --node_rank=0 --master_addr=<Master node IP address> --master_port=<Master node port number> train_gpu.py

On the second machine: python -m torch.distributed.launch --nproc_per_node=1 --nnodes=2 --node_rank=1 --master_addr=<Master node IP address> --master_port=<Master node port number> train_gpu.py

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}
}

@misc{darcet2023vitneedreg,
  title={Vision Transformers Need Registers},
  author={Darcet, Timothée and Oquab, Maxime and Mairal, Julien and Bojanowski, Piotr},
  journal={arXiv:2309.16588},
  year={2023}
}

jiaowoguanren0615 / DINOV2-Pytorch