xiangzhang1015
commented
1 year ago Hi, thanks for the questions. I'll check them later.
To keep you posted, we have updated the TFC implementation. Please check more details in the Updates on Jan 2023 section of the repo readme. In summary:
- Fixed bugs, cleaned the codes, and added comments for better understanding.
- For the contrastive encoders (in both time and frequency domains), we replaced the 3 layers of CNN blocks with 2 layers of Transformer.
- For the downstream classifier, we added a KNN classifier in parallel with the original MLP (2-layer) classifier.
- Shared more ideas that may improve the TF-C framework in follow-up works.
I have been unable to reproduce the results of the KNN (K=2) baseline on the provided datasets, except for the Epilepsy dataset where I was able to get the same score as reported in the paper. However, the performance on other datasets is significantly different. For example, in the "EMG" dataset, I am getting an accuracy score of 0.122 while the paper reports a score of 0.439.
The information provided about the HAR dataset is incorrect. The paper mentions that the dataset contains 9 channels, but upon inspection I have found that it only contains 3 channels. This discrepancy needs to be addressed.
Could you please provide more information about how the KNN baseline was run for multivariate time series data, and also correct the information about the number of channels in the HAR dataset? My goal was simply to do a quick development run, so any assistance in reproducing the results would be greatly appreciated. Thank you.