eeyhsong
commented
1 year ago Hello @ @comojin1994, sorry for the late reply.
- Yes the input uses one-second data with a sample rate of 200 Hz.
- I use 5-fold cross-validation for SEED dataset [1], instead of a 9:6 manner. Both of the two evaluation ways are widely used for this dataset. The 15 sessions of SEED were stimulated with different movie clips, and I think there may be more significant differences in the emotion produced by different sessions.
[1] Zheng, Wei-Long, Jia-Yi Zhu, and Bao-Liang Lu. "Identifying stable patterns over time for emotion recognition from EEG." IEEE Transactions on Affective Computing 10.3 (2017): 417-429.
- Yes, I averaged the results of 45 sessions to get the final results.
- I will organize the code and release it later.
- I think it is of interest to use this way of connecting convolution and self-attention to be as a new backbone for EEG analysis.
Best wishes. 🤝
comojin1994
384068026
Hi,
I'm impressed with your paper.
After reading the paper, I got some questions about using and evaluating the SEED dataset.
In the paper, you mentioned that each session contains 3394 trials, segmented from the original data using a non-overlapped one-second time window. However, I want to know more details about dataset settings.
Q1. Is the input shape of the data (62 x 200, # of channels x one-second data with sampling rate of 200 Hz)?
Q2. The number of subjects is 15, with three sessions for each subject. In addition, each session consists of 15 clips. If you follow the train&test setting in Zheng et al., Did you also split the train set as the first 9 data and the test set as the last 6 data? If this is not the case, could you explain more about this?
Q3. If Q2 is right, did you average results from 45 sessions (15 subjects x 3 sessions) for calculating results?
Q4. If it's alright with you, could you share your preprocessing code about the SEED dataset?
Thank you for sharing your excellent research.