hi, can not get result from paper, CIFAR10 IPC=10. acc of your paper is 53.5,but I only get 49.97

[Zhongwei-Luo]

======= FKD: dataset info ====== 2 path: /home/zhanglf/lzw/code/GVBSM_spkd/Branch_CIFAR_10/relabel/FKD_cutmix_fp16FKD_IPC_10 3 num img: 100 4 batch size: 100 5 max epoch: 1000 6 ================================ 7 Files already downloaded and verified 8 load data successfully 9 => loading student model 'ResNet18'

Epoch: 999 TRAIN Iter 999: lr = 0.000000, loss = 0.221218, Top-1 err = 20.000000, Top-5 err = 4.000000, train_time = 19.630981 TEST Iter 999: loss = 1.362584, Top-1 err = 50.030000, Top-5 err = 5.950000, val_time = 98.427372 wandb: wandb: wandb: Run history: wandb: train/Top1 ▃▂▆▅▆▄▂▆▅▇▂▃▂▆█▁▆▅█▅██▆▇▄▆▆▇▁▇▅▆▇▆▇▇▅▇█▂ wandb: train/Top5 ▅▄▇▆▇▅▃▆▇▇▄▅▂▇█▁█▆█▆██▇█▆▇▇▇▂█▆▇▇▇█▇▆▇█▂ wandb: train/epoch ▁▁▁▂▂▂▂▂▂▃▃▃▃▃▄▄▄▄▄▄▅▅▅▅▅▅▆▆▆▆▆▇▇▇▇▇▇███ wandb: train/loss █▅▃▃▃▃▄▃▃▂▃▂▃▂▂▄▂▂▁▂▁▁▂▁▂▂▁▁▃▁▂▂▁▂▁▁▂▁▁▃ wandb: train/lr ███████▇▇▇▇▇▆▆▆▆▆▅▅▅▄▄▄▄▃▃▃▃▂▂▂▂▂▁▁▁▁▁▁▁ wandb: val/epoch ▁▁▁▁▂▂▂▂▂▃▃▃▃▃▃▄▄▄▄▄▅▅▅▅▅▅▆▆▆▆▆▇▇▇▇▇▇███ wandb: val/loss █▃▂▂▂▂▁▂▂▁▁▁▂▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁ wandb: val/top1 ▁▄▅▆▆▆▇▇▆▇▇▇▆▇▇▇▇█▇█████▇███████████████ wandb: val/top5 ▁▆▇▇▇▇█▇█▇██▇███████████████████████████ wandb: wandb: Run summary: wandb: train/Top1 80.0 wandb: train/Top5 96.0 wandb: train/epoch 999 wandb: train/loss 0.22122 wandb: train/lr 0.0 wandb: val/epoch 999 wandb: val/loss 1.36258 wandb: val/top1 49.97 wandb: val/top5 94.05 wandb: wandb: 🚀 View run playful-sun-9 at: https://wandb.ai/485434977-shang-hai-jiao-tong/final_rn18_fkd/runs/9d0uywbl wandb: ⭐️ View project at: https://wandb.ai/485434977-shang-hai-jiao-tong/final_rn18_fkd wandb: Synced 6 W&B file(s), 0 media file(s), 0 artifact file(s) and 0 other file(s) wandb: Find logs at: ./wandb/run-20240903_181225-9d0uywbl/logs

[Zhongwei-Luo]

val/top1 49.97

[Zhongwei-Luo]

hi

[shaoshitong]

Thank you for your interest in our work! I apologize for taking so long to get back to you, as I have been very busy for a while now. This performance should be reproducible, perhaps because there are fluctuations on CIFAR-10 causing some differences in performance. If you really can't reproduce it, we give some better quality hyperparameter configurations on workEDC that you can try:

• batch szie=25
• scheduler=multistep, adjust the learning rate down to the original 0.5 at epochs [800,900,950], respectively
• ce-weight changed to 0.025

These if still not work, you can refer to the document direct_train using direct_train form (which introduces RandAugmentation for better results). In addition, we have designed Soft Category-Aware Matching and Flatness Regularization in EDC, which might be helpful for further performance improvement.