Can you provide training DVS-CIFAR10 datasets hyperparameters?

[shenhaibo123]

[fangwei123456]

You can get them at https://github.com/fangwei123456/Spike-Element-Wise-ResNet/blob/main/origin_logs/cifar10dvs/SEWResNet_ADD_T_16_T_train_None_SGD_lr_0.01_CosALR_64_amp/args.txt

This issue is also helpful: https://github.com/fangwei123456/Spike-Element-Wise-ResNet/issues/1

[shenhaibo123]

thank you

[fangwei123456]

This is the loris: https://github.com/neuromorphic-paris/loris

If you want to use the old version of SJ, you need to install it. The new version does not need it.

I recommend to use the new version of SJ to avoid the cext neuron problem (refer to this https://github.com/fangwei123456/spikingjelly/issues/46). In the new version, we use cupy to implement CUDA neuron, which avoids the compiling error of cext neuron that makes troubles to users.

[shenhaibo123]

Thank you very much for your patient answer. Just now, I had A problem with Loris, and after I solved it, I deleted the question just now (I felt A little stupid). I tried to install the latest version of SJ through “pip install SJ or git clone&& cd&& python setup.py install” , but I didn't skip the problem of Loris. I just installed Loris by switching versions and compiled this sentence. I continue to try to run SEW on my computer

[fangwei123456]

Try to run pip install spikingjelly -U.

[shenhaibo123]

Thank you for your kindly help, there may be some thing wrong, the best score I got is 72.5(the score is 74.4 in your work). I use the latest code and SJ, since the cext is not support as you mentioned here(https://github.com/fangwei123456/Spike-Element-Wise-ResNet/issues/1#issuecomment-1041061312), I replace the cext.neuron.MultiStepParametricLIFNode with clock_driven.neuron.MultiStepLIFNode as you did in (https://github.com/fangwei123456/Spike-Element-Wise-ResNet/files/8076489/dvsgesture.zip). And my parameters are the same with you. I copy the latest epoch's message: Namespace(T=16, T_max=64, T_train=None, amp=True, b=16, cnf='ADD', data_dir='/home/shb/datasets/CIFA10DVS', device='cuda:0', dts_cache='./dts_cache', epochs=64, gamma=0.1, j=4, lr=0.01, lr_scheduler='CosALR', model='SEWResNet', momentum=0.9, opt='SGD', out_dir='./logs', resume=None, step_size=32) ./logs/SEWResNet_ADD_T_16_T_train_None_SGD_lr_0.01_CosALR_64_amp epoch=63, train_loss=0.0028808565140831088, train_acc=0.9997775800711743, test_loss=1.3776003908813, test_acc=0.719, max_test_acc=0.725, total_time=148.61892414093018, escape_time=2022-03-12 00:25:44

[fangwei123456]

I just re-train this network with the current version of SJ and get 73.8 acc1. I still use the PLIF neuron. I think the accuracy around ±74 is enough.

python train.py -T 16 -data_dir /datasets/CIFAR10DVS/ -amp -lr 0.01 -cnf ADD -model SEWResNet -j 8
...
epoch=63, train_loss=0.0034258745053909003, train_acc=0.999443950177936, test_loss=1.4653774447441101, test_acc=0.722, max_test_acc=0.738, total_time=200.12506437301636, escape_time=2022-03-12 16:55:03

Here are logs and codes: cifar10dvs.zip

I notice that the origin version of this repo has some stochastic behaviors, which may cause some troubles for reproducing the idential results. Refer to the bug found at 2021-12-10 in bugs.md. These stochastic behaviors are even not controled by random seeds.

In the current version of SJ, you can try different random seeds and may get higher accuracy.