abdulfatir
commented
3 months ago @arittenbach The current Chronos models were trained with a context length of upto 512, so when you pass time series of length greater than that only the last 512 steps are actually used. See this line: https://github.com/amazon-science/chronos-forecasting/blob/main/src/chronos/chronos.py#L140
One potential way of extracting embeddings for a long time series is to feed sliding windows into the model.
I'm interested in using embeddings generated by Chronos for training a downstream anomaly detection model. For these models, if I use your sample example for generation of embeddings, I get 144 embedding vectors, which is the same length of time series in the example you provide. However, with my test data case, I have a time series of length 300000, and when I run that through your model for embedding generation I end up with 512 embedding vectors. Is there an upper bound of time series length that I should be using with this model, or is this expected output? Thanks so much!