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:

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 dinov2

For dense tasks (depth estimation and semantic segmentation), there are additional dependencies (specific versions of mmcv and mmsegmentation) which are captured in the extras dependency specifications:

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

You need to wrap up your data in a tarball file:

1. Ensure images are all in one directory

2. Create a single large tarball file that contains all images and name it pretrain_dataset.tar :

   tar -chf pretrain_dataset.tar /path/to/image/folder

1. Infer the auxiliary files pretrain_entries.npy and pretrain_file_indices.npy :

   python scripts/infer_entries.py \
       --tarball_path /path/to/pretrain_dataset.tar \
       --output_root /path/to/output/folder \
       --name pretrain

   The pretrain_entries.npy file will record:

   • a dummy class index (we set it to 0 for all images since we’re not using classes)
   • a unique filename index for each image
   • the start and end offsets of each image within the tarball file

   The pretrain_file_indices.npy file consists in a dictionnary mapping filename index to corresponding filename.

2. Dump pretrain_dataset.tar, pretrain_entries.npy and pretrain_file_indices.npy in a common folder (e.g. /root/data)

You may not want to use all the patches of a cohort, but only a subset of them (e.g. the cohort comes with a train/tune/test split and you only want to use the patches belonging to slides in the train partition).

Then, follow these simple steps:

1. Dump the image filenames (e.g. patch1.jpg) of the subset of interest in a .txt file (e.g. {subset}.txt)

2. Infer the corresponding auxiliary files pretrain_entries_{subset}.npy

   python scripts/infer_entries.py \
     --tarball_path /path/to/pretrain_dataset.tar \
     --output_root /path/to/output/folder \
     --keep /path/to/{subset}.txt \
     --name pretrain \
     --suffix {subset}

   The pretrain_entries_{subset}.npy file will record:

   • a dummy class index (we set it to 0 for all images since we’re not using classes)
   • a unique filename index for each image listed in {subset}.txt
   • the start and end offsets of each image within the tarball file

   A generic pretrain_file_indices.npy file will be saved the first time you run this command.
   It consists in a dictionnary mapping filename index to corresponding filename for the entire tarball file.

3. Dump pretrain_dataset.tar, pretrain_entries_{subset}.npy and pretrain_file_indices.npy in a common folder (e.g. /root/data)

This section describes the steps to follow in case you want to run tuning on a downstream task dataset with patch-level labels.

1. Create a .csv file containing downstream patches' filenames and labels:

   filename,label
   downstream_patch_1.jpg,3
   downstream_patch_2.jpg,1
   ...
   

2. Create a single tarball file that contains all downstream tuning patches and name it downstream_dataset.tar

   tar -chf downstream_dataset.tar /path/to/downstream/dataset/image/folder

3. Infer the auxiliary files query_entries.npy and query_file_indices.npy :

   python3 scripts/infer_entries.py \
     --tarball_path /path/to/downstream_dataset.tar \
     --output_root /path/to/output/folder \
     --csv /path/to/csv/file.csv \
     --keep /path/to/output/query.txt \
     --prefix query
   

   /path/to/csv/file.csv should point to the .csv file created in step 1. just above
   /path/to/output/query.txt should contain the list of filnames for the patches in the query subset of the downstream dataset.

4. Infer the auxiliary file test_entries.npy and test_file_indices.npy:

   python3 scripts/infer_entries.py \
     --tarball_path /path/to/downstream_dataset.tar \
     --output_root /path/to/output/folder \
     --csv /path/to/csv/file.csv \
     --keep /path/to/output/test.txt \
     --prefix test
   

   /path/to/csv/file.csv should point to the .csv file created in step 1. just above
   /path/to/output/test.txt should contain the list of filnames for the patches in the test subset of the downstream dataset.

5. dump the .tar file and the .npy files in a common folder (e.g. /root/data)

⚠️ To execute the commands provided in this section, make sure the dinov2 package is included in the Python module search path:

export PYTHONPATH="${PYTHONPATH}:/path/to/your/dinov2"

Update dinov2/configs/train/vitl14.yaml if you want to change some parameters (e.g. enabling early stopping).
Then run:

python -m torch.distributed.run --nproc_per_node=gpu dinov2/train/train.py \
    --config-file dinov2/configs/train/vitl14.yaml \
    train.dataset_path=Pathology:root={path/to/tarball/root}:extra={path/to/entry/root}:subset={subset}

Replace {path/to/data/root} with the root folder where tarballs are saved, and {path/to/entry/root} with the root folder where numpy entry files are saved (e.g. Pathology:root=/root/data:extra=/root/data).
Leave out :subset={subset} if you didn't restrict the dataset to a specific subset when preparing data.
Otherwise, replace {subset} with the suffix you chose for --suffix in data preparation (e.g. Pathology:root=/root/data:extra=/root/data:subset=train).

In case you want to run downstream tuning, make sure to update the following two parameters in your config:

tune:
  query_dataset_path: KNN:root={path/to/data/root}:extra={path/to/entry/root}:split=query
  test_dataset_path: KNN:root={path/to/data/root}:extra={path/to/entry/root}:split=test

Replace {path/to/data/root} with the folder where you dumped the downstream .tar files. Replace {path/to/entry/root} with the folder where you dumped the downstream .npy entry files.