Yonghao Song

抱歉抱歉，比较粗糙，下次一定！

Hello @pingwes, You need to train the model with your own or public datasets and then save the model weight for inference.

HI, @toncho11, Thanks for your help pointing out the issue. That may be because the BCI competition iv 2a is from the competition, which means there is just ground truth...

Hello! @rmib200, I calculated CAM for each EEG trial and chose the trials in one category for the mean CAM. 🤝

Hello @YamengGu, We provide a data shape description [here](https://github.com/eeyhsong/EEG-Conformer/blob/f17667d33671725666437b3e447e4363de4a3bce/conformer.py#L312C1-L312C77). You need to make the model input as (trials, 1, channels, time samples). API version with detailed parameter settings is available...

example data [here](https://drive.google.com/drive/folders/1a5JpW-UzZ9ASSFXFNu0wlFFTS4H2_NVv?usp=drive_link).

@rmib200 Hello, so sorry for the late reply. The `all_cam` has been obtained in the middle of the code. Then we can get the cam results corresponding to different categories...

hello @Joscelin-666, Is that 600G on GPU? You may use larger kernel of convolution module to reduce the scale of the data and also capture better features for Transformer module.

Thanks for your kindness @xuxiran 😉 The input to the net is like (batch, 1, channel, samples), as which we could arrange the data.