xuhongzuo / couta Goto Github PK

Hi, I would like to ask how to retrain this model. I tried to modify the code, but no matter how I modified it, the results were always the same as in the paper and did not change.

Hello, @xuhongzuo. This is a great work. I am interested with data perturbation(which also called data augmentation in other paper). Data augmentation is a popular work, and this technique is often used in contrastive learning. I noticed that this is used in recent time series anomaly detection work, such as COCA, TFAD. In the paper, you use several simple transformations to generate abnormal samples. Where is the code for this processing?

@xuhongzuo . I want to use pd.DataFrame to save loss, loss_oc and val_loss in each epoch, and save it as a csv file. Could you add the code for this function?

Hi, I have a problem during training. When I removed the backpropagation and parameter updates during training, the model was still able to train correctly and produced similar results to the paper.

Hello,
First, thank you for your work, I've been using USAD so far but your solution seems to give similar or even better results on my data, while being easier to tune/train

I'm not really familar with models built that way, so I'm kinda stuck on one point, if you could guide me I'd be grateful:
I trained my model on colab, but need to make it available on another computer (inference on CPU only) for experimentation purpose.

Here is the struggle : I Can't find the proper way to save the model to be used elsewhere.

What I tried :

• dill/pickle/torch.save the trained model object, rearrange functions,... but there's always at least one of the dependencies that cannot be interpreted. I know it's not the best pratice to do so anyway
• Just saved the 'net' state_dict, recreate the COUTA instance and load dict but too much things are done in fit(), so I can't reload weights if the model is not fitted before

Any help will be appreciated, thank you in advance