The config for cifar10 in Paper

[FengLiXue]

Hi, First of all thanks for this great work. I have tried all the configs under the conf/cifar10 folder, but none of them can meet the accuracy in the paper. Could you provide detailed configurations that can meet the accuracy in the paper? Thank you very much!

[jkhu29]

https://github.com/BICLab/Spike-Driven-Transformer/blob/main/conf/cifar10/2_512_300E_t4.yml is the conf we used.

If the model fails to achieve an accuracy of 95% on the cifar10 dataset after you train with this config, I would recommend you try to extend the number of layers to 4 (but you don't actually have to do that, maybe it's a random number seed problem, or maybe it's an issue in the hardware and software environment, such as an old CUDA version)

[FengLiXue]

Thanks for reply. I tried conf/cifar10/2_256_300E_t4.yml, 2_512_300E_t4.yml and 2_512_300E_t4_TET.yml. I also extend the number of layers to 4, but all the accuracy is about 85%, there is a big gap between 95%, it doesn't seem like a random seed problem. The GPU I used is 3090, and the CUDA version is 11.6, pytorch version is 1.13.0. I don't know where I didn't do it right. If you could provide help, I would be very grateful!

[jkhu29]

Maybe you can provide us with your tensorboard file and we can analyze your loss or accuracy together.

[jkhu29]

You can see that the model accuracy has reached 86.35 at epoch 27, which is the result of the two RTX3090 I just used to run this git.

[jkhu29]

The launch command is

CUDA_VISIBLE_DEVICES=0,1 /usr/bin/python3 -m torch.distributed.launch --nproc_per_node=2 --master_port 29501 train.py -c conf/cifar10/2_512_300E_t4.yml --model sdt --spike-mode lif

[FengLiXue]

@jkhu29 Firstly, I am very grateful for the work you have done for my problem. I have put the training log here. It would be great if you could help me take a look. As the logs, when the accuracy reaches 84%， the model accuracy will improve very slowly, this makes me very confused. Thank you very much again!

[jkhu29]

The batch-size you used is 64, CIFAR10 has 50000 training files and 10000 testing files, so the log should be [0/780], not the 3298. I guess there is something wrong with your data, please follow the Pytorch website documentation to configure your data.

[FengLiXue]

Thank you very much for pointing out my mistake. I did indeed make a mistake with the dataset, and I am very sorry for taking up so much of your time. Once again, I would like to express my gratitude for your work

[jkhu29]

Great, I hope it works for you. If your problem has been solved, please close this issue, you can reopen it if you have any further problems.