search

putshua / ANN_SNN_QCFS

Code for paper "Optimal ANN-SNN conversion for high-accuracy and ultra-low-latency spiking neural networks"
24 stars 6 forks source link

unable to reproduce the result on CIFAR-100 with the same settings in the paper #1

Open Berumotto1 opened 1 year ago

Berumotto1 commented 1 year ago

In the experiment of VGG16 on CIFAR100, I set the l=4, init_threshold=4, seed=42, the acc of SNN at T=2 is 0.0100 and the acc of SNN at T=64 is 0.7534(0.7705 in the paper). In the experiment of ResNet20 on CIFAR100, I set the l=8, init_threshold=4, seed=42, the acc of SNN at T=64 is 0.6890(0.7055 in the paper). Thank you for providing the code, expecting for your reply.

putshua commented 1 year ago

Please refer to the latest version.

bjourne commented 1 month ago

Hello!

Fwiw, I'm am able to reproduce your results on CIFAR100 with VGG16 and L=8:

SRC ANN T=1 T=2 T=4 T=8 T=16 T=32 T=64 T=128
Mine 77.15 43.08 52.73 63.49 71.77 75.71 76.87 77.05 77.10
Google Drive 77.41 35.38 52.71 66.04 70.75 73.54 74.43 74.34 74.41
Paper - 44.98 52.46 62.09 70.71 74.83 76.41 76.73 76.73

Mine refers to the model I trained myself, Google Drive to your pretrained model, and Paper to the figures in the paper. I even get slightly better performance for some values of T. Only difference is that I train with LR=0.1 instead of 0.05.