Open sauravtii opened 2 years ago
Hi, I am trying out this code (https://github.com/fangwei123456/spikingjelly/blob/master/spikingjelly/activation_based/ann2snn/examples/resnet18_cifar10.py). The only difference is I am using resnet-34 from this file (https://github.com/fangwei123456/spikingjelly/blob/master/spikingjelly/activation_based/ann2snn/sample_models/cifar10_resnet.py). I trained it, converted it to SNN and also simulated it. The accuracy that I got after running the resnet-18 architecture was 94.1 % (shown below).
ANN accuracy: 100%|██████████| 200/200 [00:02<00:00, 96.58it/s] Validating Accuracy: 0.946 Converting... 100%|██████████| 500/500 [00:09<00:00, 52.04it/s] SNN accuracy: 100%|██████████| 200/200 [05:27<00:00, 1.64s/it] Validating Accuracy: 0.941
For resnet-34 I am getting the following accuracies after training it for 150 epochs, converting it to SNN and simulating it.
Epoch: 149 100%|██████████| 500/500 [00:07<00:00, 65.69it/s] Validating Accuracy: 0.965 100%|██████████| 200/200 [00:01<00:00, 106.85it/s] ANN Validating Accuracy: 0.8307 --------------------------------------------- Converting using MaxNorm 100%|██████████| 500/500 [00:08<00:00, 57.73it/s] Simulating... 100%|██████████| 200/200 [17:48<00:00, 5.34s/it] SNN accuracy (simulation 700 time-steps): 0.8248 --------------------------------------------- Converting using RobustNorm 100%|██████████| 500/500 [05:52<00:00, 1.42it/s] Simulating... 100%|██████████| 200/200 [18:01<00:00, 5.41s/it] SNN accuracy (simulation 700 time-steps): 0.8073 --------------------------------------------- Converting using 1/2 max(activation) as scales... 100%|██████████| 500/500 [00:08<00:00, 56.58it/s] Simulating... 100%|██████████| 200/200 [18:03<00:00, 5.42s/it] SNN accuracy (simulation 700 time-steps): 0.8160
Can anyone please tell me how can I get the same accuracy for resnet-34 and also for other resnet architectures as mentioned here ?
Just wanted to follow up.
Hi, I am trying out this code (https://github.com/fangwei123456/spikingjelly/blob/master/spikingjelly/activation_based/ann2snn/examples/resnet18_cifar10.py). The only difference is I am using resnet-34 from this file (https://github.com/fangwei123456/spikingjelly/blob/master/spikingjelly/activation_based/ann2snn/sample_models/cifar10_resnet.py). I trained it, converted it to SNN and also simulated it. The accuracy that I got after running the resnet-18 architecture was 94.1 % (shown below).
For resnet-34 I am getting the following accuracies after training it for 150 epochs, converting it to SNN and simulating it.
Can anyone please tell me how can I get the same accuracy for resnet-34 and also for other resnet architectures as mentioned here ?