search

emadeldeen24 / TSLANet

[ICML 2024] A novel, efficient approach combining convolutional operations with adaptive spectral analysis as a foundation model for different time series tasks
https://arxiv.org/abs/2404.08472
MIT License
115 stars 21 forks source link

Unused nn.Linear in classification model and paper inaccuracy? #8

Open Snonky opened 2 weeks ago

Snonky commented 2 weeks ago

Hi, great work, thank you!

I found an orphaned nn.Linear module that was presumably replaced by the PatchEmbed layer in the classification model. I believe this line should be removed.

https://github.com/emadeldeen24/TSLANet/blob/ca0e88416d3ae49fd50e399c44ae94868378a94d/Classification/TSLANet_classification.py#L158

Slightly related: in section 3.3 of the paper it is stated that Pš‘– ā†’ P'š‘– āˆˆ ā„CāØÆš‘' with C being the number of channels. As far as I understand š‘' is the embedding size (--emb_dim in the code) and that is used to embed all dimensions into one so it should be P'š‘– āˆˆ ā„š‘' dropping the channel dimension. I apologize if I missed something here that proves my remark as wrong.

luojike2333 commented 4 days ago

Good question! Have you tried conducting experiments to explore the differences between using nn.Linear and not using it? Whether there is a difference in the results of the experiment?

luojike2333 commented 1 day ago

Good question! Have you tried conducting experiments to explore the differences between using nn.Linear and not using it? Whether there is a difference in the results of the experiment?