Fail to Reproduce SpikFormer-8-512 on Imagenet

[CatherineCloud]

Hi dear authors,

Thank you guys for your amazing work. May I ask if there is anything important that requires modification based on the current code uploaded to github? I was trying to reproduce the result for Spikformer-8-512 on Imagenet. However, the result I got was quite different from the one shown in the paper.


When it comes to epoch 25, the top1-acc shown in the paper is clearly over 50%, while my reproduced result is barely over 40%. I used the exactly same code as in this github, except I used a batch size of 24 due to my GPU limitation.

Please enlight me. Thank you so much!

[ZK-Zhou]

Please provide the curve of the full training 300 epoches, because the training process will be different with different hyperparameters, rather than just looking at the partial convergence curve

[CatherineCloud]

The hyperparameters are not changed as the one shown in this github repo.
Additionally, may I ask how long did you guys take to train thie model on ImageNet?

[ZK-Zhou]

Hi, batch size is one of the most important hyperparameters 。We use 8 Nvidia V100 for 8 days.

[warming151]

Hi. May I ask which PyTorch version you installed? I couldn't install pytorch==1.10.0+cu111... Thanks!