a question

[nince9]

Hello dude, I am a beginner attempting to reproduce your article. However, I have found that the accuracy and training curves on Wandb are very poor. Could you please advise if this issue is due to my usage or evaluation process?
with:
python train_test.py -g 0,1,2 --num_nodes 1 -n 4 -b 512 -e 2000 \ --data_dir /data/datasets/imagenet/ --pin_memory --wandb_project_name fasternet \ --model_ckpt_dir ./model_ckpt/$(date +'%Y%m%d_%H%M%S') --cfg cfg/fasternet_t0.yaml
the results:

thanks for your reply

[JierunChen]

Hi, you could try enabling the clip_grad with proper value (e.g., 1) in the cfg file.