MTAD-GAT #9
MTAD-GAT #9
Conversation
gonzachiar
commented
Jun 23, 2024
- Add MTAD-GAT implementation (training and inference).
- Inference and metrics are done in a terrible way, this will change in the future.
- Added new function to load datasets easier
| X, | ||
| Y, | ||
| test_size=test_size, | ||
| shuffle=shuffle, |
There was a problem hiding this comment.
Hm I think we shouldn't be doing shuffling nor stratifying in time-series
There was a problem hiding this comment.
Be careful, it's done in several places
| class SlidingWindowDataset(Dataset): | ||
| def __init__( | ||
| self, | ||
| data: torch.Tensor, # TODO: Check if a tensor is a numpy ArrayLike |
There was a problem hiding this comment.
Seem to be:
However, we can treat PyTorch tensors as NumPy arrays without the need for explicit conversion:
https://numpy.org/doc/stable/user/basics.interoperability.html
|
|
||
|
|
||
| # TODO: Standarize | ||
| def adjust_predicts(score, label, threshold, pred=None, calc_latency=False): |
There was a problem hiding this comment.
Wow this function sucks. I would try to improve a little bit that for with a million of if's if possible. I don't like that the last return returns just a value instead of a tuple, feels like a trouble maker. I'd add return predict, None like in the first return statement
There was a problem hiding this comment.
I don't know about this. It just feels like a lot of logic (that we don't really understand) to run only one evaluation method. Maybe we can remove that evaluation method for now and if it's really useful we re-consider it? The pot_eval
| from models.mtad_gat.model import MTAD_GAT | ||
| from models.mtad_gat.utils import adjust_anomaly_scores | ||
|
|
||
| # TODO: |