[CIFAR10] Fine-tuning accuracy is far below the reported value (56.9 % vs 81.0 %)

[KT-muscle]

Overview

Thank you for releasing the code!
Following the README, I fine-tuned DeiT-Small (DINO-pretrained) using the subset
data_selection/subsets/CIFAR10/Density_250_from_ActiveFT50.json
(i.e., 250 labelled images = 0.5 % of CIFAR-10).
However, I obtained Top-1 = 56.93 %, which is much lower than the 81.0 ± 1.2 % reported in the paper.

Reproduction commands

# -------- fine-tuning --------
python -m torch.distributed.launch --nproc_per_node=1 main.py \
  --data-set CIFAR10SUBSET \
  --subset_ids ../data_selection/subsets/CIFAR10/Density_250_from_ActiveFT50.json \
  --resume dino_deitsmall16_pretrain.pth \
  --output_dir Outputs/C10_250 \
  --epochs 1000 \
  --batch-size 512\
  --lr 2.5e-4 \
  --eval_interval 50 \
  --clip-grad 2.0

# -------- evaluation --------
python main.py \
  --data-set CIFAR10 \
  --data-path data \
  --batch-size 512 \
  --resume Outputs/C10_250/best_checkpoint.pth \
  --eval \
  --save_metrics Outputs/C10_250/metrics.json