-
Notifications
You must be signed in to change notification settings - Fork 25
Open
Description
Lines 11 to 19 in 953b833
| def FFT_for_Period(x, k=2): | |
| # [B, T, C] | |
| xf = torch.fft.rfft(x, dim=1) | |
| frequency_list = abs(xf).mean(0).mean(-1) | |
| frequency_list[0] = 0 | |
| _, top_list = torch.topk(frequency_list, k) | |
| top_list = top_list.detach().cpu().numpy() | |
| period = x.shape[1] // top_list | |
| return period, abs(xf).mean(-1)[:, top_list] |
Thanks for this great work! However, the rationale behind the implementation of FFT_for_Period may not be clear.
- In my opinion, the
FFT_for_Periodfunction is used for frequency selection, which is a necessary preprocessing step for the multi-scale modeling described in the paper. It returns the scale sizes with corresponding top-k largeset amp.. - However, in line 14, the frequency selection is based on the TS averaging across batches, which may not be appropriate during testing. How would model performance be influenced if the calculation were done on a per-sample manner?
Metadata
Metadata
Assignees
Labels
No labels