search

zbs881314 / Temporal-Coded-Deep-SNN

MIT License
27 stars 7 forks source link

Replicating CIFAR10 results on your paper #4

Open TakeoHosomi opened 1 year ago

TakeoHosomi commented 1 year ago

Thank you for sharing your code. I'm interested in your work ang trying to replicate your CIFAR10 results. However I have two issues.

  1. I only got 61% when I run TrainCifar10_S.py without loading weights and run TestCifar10_S.py, though pre-trained weights shows 64% when I run TestCifar10_S.py. Could you let me know how to generate your pre-trained weights from scratch?

  2. In your paper, you wrote "Therefore, we limited the maximum allowed row-normalized Frobenius norm of the gradient of each weight matrix to 10." at page 5. Could you let me know which part of the code implement this feature.

TakeoHosomi commented 1 year ago

I have modified some and got 69% on TestCifar_S.py:

  1. use lr_decay as TrainCifar_L.py
  2. use Cifar10_with_data_augmentation.CFAR10() as dataloader
  3. disable whiten option at Cifar10_with_data_augmentation.CFAR10