Could you please share the random seed for training and label for BCI data E file? about eeg-transformer

[110lmy]

Excuse me, how does the part of location code work? Do I take sines and cosines?

[eeyhsong]

@zewailaa hello, you may see the randomly generated seed in the main function at the end of the code.

[eeyhsong]

@110lmy You mean position embedding? We don't use this design in the code.

[110lmy]

It's not positional embedding, it's positional coding. Wasn't position coding used before compression slices in the time conversion section?

[eeyhsong]

@110lmy Sorry for the misunderstanding. In my provious experiments, the position embedding strategies as used in images performed not well for EEG signals. So I chose to use one convolutional layer to capture the position information.

[110lmy]

卷积层获取位置信息是怎么获取的？我还想问一下，为什么在空间变换使用单头注意力，但是在时间变换使用多头注意力？谢谢回复 | | 刘梦园 | | ***@***.*** | ---- Replied Message ---- | From | ***@***.***> | | Date | 11/7/2022 11:32 | | To | ***@***.***> | | Cc | ***@***.***> , ***@***.***> | | Subject | Re: [eeyhsong/EEG-Transformer] Could you please share the random seed for training and label for BCI data E file? (Issue #6) | @110lmy Sorry for the misunderstanding. In my provious experiments, the position embedding strategies as used in images performed not well for EEG signals. So I chose to use one convolutional layer to capture the position information. — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you were mentioned.Message ID: ***@***.***>

[eeyhsong]

@110lmy hi, lmy

• In my view, position information in EEG is still a kind of local correlation, which can be encapsulated by convolution.
• You could also use multi-head in channel attention. That is a performance enhancing strategy, but the interpretability for EEG needs to be further explored.
• Besides, I found too long data is not easy to caluculate the attention weight well.
• At last, please wait for the next version.