kthyeon / fedsup Goto Github PK

First of all, thanking you for your valueable work and for providing the code. When I try to reproduce the experimental results of E-Fedsup in Table 1 in the paper, the training process of the model dose not seem to work as expected. The performance of the model is stable at 1% accuracy, while the loss value keeps oscillating. The final performance did not behave as demonstrated in the paper. Could you help me to tell the cause of this result? Or tell me how to reproduce the experimental results of Table 1 in the paper. Thank you for your reply.
The trainning command I use:

python main.py --shard_per_user 10 --dataset cifar100 --label_smoothing 0.0 \
                          --gpu 0 --frac 0.1 --local_ep 5 --momentum 0.9 \
                          --inplace_distill False --num_arch_training 1 \
                          --drop_connect 0.0 --dropout 0.0 --results_save "vanilla" \
                          --diri "False" --config-file config/efficient_fedsup.yml

The config file efficient_fedsup.yml has not been modified.

part of the training results are shown in the figure below: