backbone

[zzj2404]

thank you for your time series work.
for transformer backbone, my SleepEEG2Epilepsy reproduce result is close to paper，but others i can't get a good result.
I refered to the paper and simCLR baseline code, try to build and train Resnet backbone but result is bad.
Could you please give a detail structure of Resnet backbone?
Thank you!

[yuyunannan]

I have the same problem

[findalexli]

Can you share how you reproduce the SleepEEG2Epilepsy paper result?

[yuyunannan]

I'm guessing now that the author just gave the correct configuration of sleepeeg experiments?

[zzj2404]

Can you share how you reproduce the SleepEEG2Epilepsy paper result?
I'm guessing now that the author just gave the correct configuration of sleepeeg experiments?

In sleepeeg config, I set batchsize from 128 to 64; in model , set transformer encoder to 3 layers; in main, let subset = True.these are same with paper except for subset.
Then I get a result of acc=94.1944, precision=93.7874, recall=87.4317, F1=90.1626.
since using transformer will have a drop of performance, I think the reproduce is successful?
But when using 'subset=False' the result is even worse than without pretraining.

[xiaoyuan7]

I've noticed a small issue with using TransformerEncoderLayer. When the input tensor shape is [B, C, T] and TransformerEncoderLayer is applied directly, attention seems to be applied on B instead of T. I'm not entirely sure if I'm understanding this correctly, but I was wondering if the author could provide some clarification on this matter. Additionally, would it be possible for the author to submit the backbone of ResNet? Thank you

[zzj2404]

I meet with the same problem. Do you successfully reproduce SleepEEG to other dataset?

no

[rybd111]

I have the same issue. When using the sleepEEG dataset for pre-training and fine-tuning with other datasets, all categories are classified into the same class, resulting in a warning of no positive samples and low accuracy. I really hope the author can provide the data preprocessing process, preferably in code form.