search

prajwalsingh / EEGStyleGAN-ADA

Pytorch code of paper "Learning Robust Deep Visual Representations from EEG Brain Recordings". [WACV 2024]
MIT License
27 stars 4 forks source link

About EEG2Feat #14

Closed 7ASSEL closed 5 months ago

7ASSEL commented 5 months ago

Hello bro, I was training EEG2Feat-LSTM on the CVPR40 5_95Hz dataset, And after 1000+epochs, the K_means acc of the train_set was close to 1.0, but the K_means acc of validation was only 0.14+, Should I make any changes to the network or hyper-paramaters, Or should I switch to the raw CVPR40 dataset to reproduce the 0.9+ result on validation?

prajwalsingh commented 5 months ago

Hi @7ASSEL , the triplet loss based LSTM feature extraction network works for raw only. So you have to used raw for val acc. of 90+

7ASSEL commented 5 months ago

Hi @7ASSEL , the triplet loss based LSTM feature extraction network works for raw only. So you have to used raw for val acc. of 90+

Thanks a lot! It really worked. And besides the EEGClip experienment on the filtered 5-95Hz dataset mentioned in your paper, is there any other useage of the 5-95Hz dataset?

prajwalsingh commented 5 months ago

@7ASSEL , other than experiment we haven't used it for anything else. At present we are accumulating different EEG-Image dataset for diversity.