Closed dara1400 closed 3 months ago
These is are other simple tests, which model can not follow trend:
Hi Hamzeh, Thanks for your interest in MOMENT! As we have mentioned in the forecasting tutorial, MOMENT is pre-trained to reconstruct input time series. To use MOMENT for forecasting, we replace the reconstruction head with a forecasting head-- a linear layer which maps MOMENT embeddings to the forecasting horizon. This forecasting head is randomly initialized and must be fine-tuned before use. We have made this clearer in the tutorial now 🤗
While MOMENT can do zero-shot reconstruction, imputation, anomaly detection, and unsupervised representation learning for classification, it does not do zero-shot forecasting.
I would recommend fine-tuning MOMENT on your data for a few epochs. You can find example fine-tuning code in the forecasting tutorial. As you can see from the tutorial, with only 1 epoch and no hyperparameter tuning, MOMENT can get competitive forecasting results!
I'll close this issue for now, but please feel free to re-open this issue, or create another one if you face more problems! Again, thank you so much for your interest in MOMENT!
I used a dataset of a demand of electric load for one month. The step is one hour. The values are normalized. I feed 512 point to the model and forecasting horizon is 24. The output is not good.