Simple questions about dataset preparation before training DINOV2

[EddieAy]

@woctezuma Sorry to bother you again. But I just wanna know how can i get correct ImageNet-like dataset

• For example,when i train dinov1, it's simple and it just uses a ImageFolder

• Below picture is the dataset preparation process in DINOV1

• And this is my dataset structure. I don't need any other adjustment to this dataset when i try to train DINOV1. Because this is what ImageFolder likes

• BUT NOW, when it comes to dinov2,it shows that:

-So do i need an /labels.txt? Like below and i don't have TEST SET

-Although i don't know what are those files under 'extra', when i follow the below code

• I got this. So how can i correct? How i get start to train dinov2?

• And this is my dataset structure

• 

Thank you @woctezuma

[XiaohuJoshua]

+1, it would be better if we can just uses a ImageFolder

[onvungocminh]

@EddieAy @XiaohuJoshua Did you solve the problem?

[HDL-YD]

I have the same doubt.

[bruce-willis]

There are several ways how you can train linear head for classification on your data:

1. Downstream-Dino-V2 repository
2. HF Transformers version of DINOv2 with the example notebook for the classification
3. Modify this repository a little bit.

I will advocate for the last approach because it contains the training of many linear heads simultaneously with different parameters (learning rate, average pooling or not and different blocks from the backbone). I will do for the imagenette dataset, but you, of course, can use whatever you want.

• Create a file with torchvision dataset as usual here dinov2/data/datasets
• Additionally implement get_targets method, which should return the number of classes in your dataset. This method is used here: https://github.com/facebookresearch/dinov2/blob/2302b6bf46953431b969155307b9bed152754069/dinov2/eval/linear.py#L499
• Here is the full source code for reference.
• Add the following line in dinov2/data/datasets/__init__.py file:

  from .imagenette import Imagenette

• Also import your dataset here https://github.com/facebookresearch/dinov2/blob/2302b6bf46953431b969155307b9bed152754069/dinov2/data/loaders.py#L13
• Add download key (or any other which you are using) to the list of available keys here https://github.com/facebookresearch/dinov2/blob/2302b6bf46953431b969155307b9bed152754069/dinov2/data/loaders.py#L52
• Add your dataset here https://github.com/facebookresearch/dinov2/blob/2302b6bf46953431b969155307b9bed152754069/dinov2/data/loaders.py#L55-L60

  elif name == "imagenette":
  class_ = Imagenette

• Launch training (don't forget to update paths or any other parameters):

  CUDA_VISIBLE_DEVICES='0' python ./dinov2/eval/linear.py --config-file ./dinov2/configs/eval/vits14_pretrain.yaml --pretrained-weights ./weights/dinov2_vits14_pretrain.pth --output-dir results/dinov2_vits14/without_rtokens/linear/imagenette --train-dataset imagenette:split=train:root=./data/imagenette:download=True --val-dataset imagenette:split=val:root=./data/imagenette --batch-size 128

Feel free to contact me if you have any issues.

[PMRS-lab]

@woctezuma Sorry to bother you again. But I just wanna know how can i get correct ImageNet-like dataset

* For example,when i train dinov1, it's simple and it just uses a ImageFolder

* Below picture is the dataset preparation process in **DINOV1**
  ![image](https://private-user-images.githubusercontent.com/101048881/240350680-cca88cae-dcb9-498a-94c7-dfe58a4f4d91.png?jwt=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJnaXRodWIuY29tIiwiYXVkIjoicmF3LmdpdGh1YnVzZXJjb250ZW50LmNvbSIsImtleSI6ImtleTUiLCJleHAiOjE3MjQ3NzQzOTIsIm5iZiI6MTcyNDc3NDA5MiwicGF0aCI6Ii8xMDEwNDg4ODEvMjQwMzUwNjgwLWNjYTg4Y2FlLWRjYjktNDk4YS05NGM3LWRmZTU4YTRmNGQ5MS5wbmc_WC1BbXotQWxnb3JpdGhtPUFXUzQtSE1BQy1TSEEyNTYmWC1BbXotQ3JlZGVudGlhbD1BS0lBVkNPRFlMU0E1M1BRSzRaQSUyRjIwMjQwODI3JTJGdXMtZWFzdC0xJTJGczMlMkZhd3M0X3JlcXVlc3QmWC1BbXotRGF0ZT0yMDI0MDgyN1QxNTU0NTJaJlgtQW16LUV4cGlyZXM9MzAwJlgtQW16LVNpZ25hdHVyZT0xNjFjMmQ5ZmE1MGE1ZjgyMTNlODdjYzNjYmMwZjczZGM4YjY4M2EwODFkYmJkNzBlNDUxMTMzZDA4MjRkZDZmJlgtQW16LVNpZ25lZEhlYWRlcnM9aG9zdCZhY3Rvcl9pZD0wJmtleV9pZD0wJnJlcG9faWQ9MCJ9.1N7YeTXofhQmPiwT7tne37p8qGNiHgKt63yBSVTRMoM)

* And this is my dataset structure. I don't need any other adjustment to this dataset when i try to train **DINOV1**. Because this is what ImageFolder likes
  ![image](https://private-user-images.githubusercontent.com/101048881/240351056-aa8cef2e-f15e-47c3-a914-c94ed77a4536.png?jwt=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJnaXRodWIuY29tIiwiYXVkIjoicmF3LmdpdGh1YnVzZXJjb250ZW50LmNvbSIsImtleSI6ImtleTUiLCJleHAiOjE3MjQ3NzQzOTIsIm5iZiI6MTcyNDc3NDA5MiwicGF0aCI6Ii8xMDEwNDg4ODEvMjQwMzUxMDU2LWFhOGNlZjJlLWYxNWUtNDdjMy1hOTE0LWM5NGVkNzdhNDUzNi5wbmc_WC1BbXotQWxnb3JpdGhtPUFXUzQtSE1BQy1TSEEyNTYmWC1BbXotQ3JlZGVudGlhbD1BS0lBVkNPRFlMU0E1M1BRSzRaQSUyRjIwMjQwODI3JTJGdXMtZWFzdC0xJTJGczMlMkZhd3M0X3JlcXVlc3QmWC1BbXotRGF0ZT0yMDI0MDgyN1QxNTU0NTJaJlgtQW16LUV4cGlyZXM9MzAwJlgtQW16LVNpZ25hdHVyZT1kOWZmNWEzMDdkYTNhZTQxMzM1Y2JkNmM4MzIyMjdhMmUxZDBkZDFkMzRiY2Q4ZTk1NmI1MTc0ZWQzZTE5ZWQ4JlgtQW16LVNpZ25lZEhlYWRlcnM9aG9zdCZhY3Rvcl9pZD0wJmtleV9pZD0wJnJlcG9faWQ9MCJ9.rIbDNA2u48IivpN0-3ixm-yJe_DdmE-ZduGW0F2dhcs)

* BUT NOW, when it comes to dinov2,it shows that:
  ![image](https://private-user-images.githubusercontent.com/101048881/240351623-54d80d18-715f-47f0-b3e1-c408157b31f7.png?jwt=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJnaXRodWIuY29tIiwiYXVkIjoicmF3LmdpdGh1YnVzZXJjb250ZW50LmNvbSIsImtleSI6ImtleTUiLCJleHAiOjE3MjQ3NzQzOTIsIm5iZiI6MTcyNDc3NDA5MiwicGF0aCI6Ii8xMDEwNDg4ODEvMjQwMzUxNjIzLTU0ZDgwZDE4LTcxNWYtNDdmMC1iM2UxLWM0MDgxNTdiMzFmNy5wbmc_WC1BbXotQWxnb3JpdGhtPUFXUzQtSE1BQy1TSEEyNTYmWC1BbXotQ3JlZGVudGlhbD1BS0lBVkNPRFlMU0E1M1BRSzRaQSUyRjIwMjQwODI3JTJGdXMtZWFzdC0xJTJGczMlMkZhd3M0X3JlcXVlc3QmWC1BbXotRGF0ZT0yMDI0MDgyN1QxNTU0NTJaJlgtQW16LUV4cGlyZXM9MzAwJlgtQW16LVNpZ25hdHVyZT01NjBiNGIxOGEwOTYzM2ZmOWQ2YWY0OTYwMGNmMThmMTQ2ZWMwYjYwMjBkMTQ4ZTNlZDg5ZmMwZmFjYTU4ZTg1JlgtQW16LVNpZ25lZEhlYWRlcnM9aG9zdCZhY3Rvcl9pZD0wJmtleV9pZD0wJnJlcG9faWQ9MCJ9.QJ4oLfV8kEbFdBRks1jcfvz4_SU45WJWTzZ6H8zxP1M)

-So do i need an /labels.txt? Like below and i don't have TEST SET

-Although i don't know what are those files under 'extra', when i follow the below code

* I got this.  So how can i correct?   How i get start to train dinov2?
  ![image](https://private-user-images.githubusercontent.com/101048881/240352539-956f6fcf-60c0-41e7-bdc2-a21cca3fe018.png?jwt=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJnaXRodWIuY29tIiwiYXVkIjoicmF3LmdpdGh1YnVzZXJjb250ZW50LmNvbSIsImtleSI6ImtleTUiLCJleHAiOjE3MjQ3NzQzOTIsIm5iZiI6MTcyNDc3NDA5MiwicGF0aCI6Ii8xMDEwNDg4ODEvMjQwMzUyNTM5LTk1NmY2ZmNmLTYwYzAtNDFlNy1iZGMyLWEyMWNjYTNmZTAxOC5wbmc_WC1BbXotQWxnb3JpdGhtPUFXUzQtSE1BQy1TSEEyNTYmWC1BbXotQ3JlZGVudGlhbD1BS0lBVkNPRFlMU0E1M1BRSzRaQSUyRjIwMjQwODI3JTJGdXMtZWFzdC0xJTJGczMlMkZhd3M0X3JlcXVlc3QmWC1BbXotRGF0ZT0yMDI0MDgyN1QxNTU0NTJaJlgtQW16LUV4cGlyZXM9MzAwJlgtQW16LVNpZ25hdHVyZT0wZDc2ODJhNDNjOGJmNzE3OGFmN2ZiZThmZGM4ZjU3MDZmNzk5Njk0OTlhNTlmZjg3NTM2YTAyNjJkOTg3MDZhJlgtQW16LVNpZ25lZEhlYWRlcnM9aG9zdCZhY3Rvcl9pZD0wJmtleV9pZD0wJnJlcG9faWQ9MCJ9.b2q1Ij5Iun9u5U9RcSJyzN8rp2iYh9B1x-LPgDZUhUo)

* And this is my dataset structure

Thank you @woctezuma

Did you solve the problem?