exps_dino

Running DINO Experiments

This folder includes code for training and evaluating DINO variants based on customizations of dino and ssl-transfer.

• Running DINO Experiments
  • Getting started
    • Data preparation
    • Environment setup
  • Running
    • Training
    • Evaluation

Getting started

Data preparation

We support training DINO on folder-based image datasets (e.g., ImageNet and rearranged LAIONet) and csv-based datasets (e.g., ImageNet-Captions and LAIONet subsets). The example scripts use the former by default.

For transfer learning evaluation, we support 12 downstream datasets, including CIFAR10/100, Aircraft, Birdsnap, Caltech101, Cars, DTD, Flowers, Food, Pets, SUN397, and VOC. Please download the datasets following the instructions in ssl-transfer, and the resulting file structure should follow details in config.py.

Environment setup

Existing environment used for SL and CLIP training should be sufficient. Please refer to the README for reference.

Running

Training

We privide example scripts to replicate our experiments in the scripts folder. This includes training DINO on ImageNet and LAIONet, with different numbers of prototypes (fixed), or different numbers of prototypes to sample from the 65536 ones when computing the loss. We train with 8 GPUs by default. Checkpoints are saved to the output folder by default.

Evaluation

After training is done, first convert the checkpoint to desiered format by running the following command:

python convert_dino_ckpt.py --model $PATH_TO_CHECKPOINT

By default, the converted checkpoint named dino.pth will be saved to the same directory as the original checkpoint. Then, run the following command to evaluate the checkpoint on downstream datasets with linear probing:

python scripts/run_linear_transfer.sh $PATH_TO_CHECKPOINT $DATASET_NAME # e.g., cifar10

The results will be saved to linear.csv in the same directory as the checkpoint file.

Note that the scripts also supports evaluating SL and CLIP checkpoints, and few-shot/fully fine-tune evaluations. Feel free to explore the code.