Why is your code execution more efficient?

[Kiros-yzl]

[Kiros-yzl]

Hello, thanks for you work! I recently executed your code and found that your TimesNet model is much faster than the code published by the original author. In the case where the batch size and sequence length are consistent, your model can achieve an epoch of 3 seconds, but the code published by the original author can only achieve an epoch of 2400 seconds.( Test data is WADI.)
May I ask where you have optimized? I am looking forward to your reply. Thank you very much.

[YYZStart]

Hello, thanks for you work! I recently executed your code and found that your TimesNet model is much faster than the code published by the original author. In the case where the batch size and sequence length are consistent, your model can achieve an epoch of 3 seconds, but the code published by the original author can only achieve an epoch of 2400 seconds.( Test data is WADI.) May I ask where you have optimized? I am looking forward to your reply. Thank you very much.

Hi!
I looked into the code and the main speed-up comes from the epoch_steps parameter:

# in BaseDeepAD / TimesNet wrapper epoch_steps = 20 # → stops after 21 mini-batches

If you want to train on the entire DataLoader, just set

epoch_steps = -1 # process all batches every epoch

[GJ-FZU]

Hello, thanks for you work! I recently executed your code and found that your TimesNet model is much faster than the code published by the original author. In the case where the batch size and sequence length are consistent, your model can achieve an epoch of 3 seconds, but the code published by the original author can only achieve an epoch of 2400 seconds.( Test data is WADI.) May I ask where you have optimized? I am looking forward to your reply. Thank you very much.

你好，我也想使用这个项目对WADI数据集进行异常检测，请问数据集要做什么预处理，是直接使用吗，方便加一个联系方式吗