Timestamps and embeddings..?

[vstenby]

Hi - completely new to this, so sorry if this is a stupid question!

I want to use chronos-2 as an encoder for some time series.

from chronos import Chronos2Pipeline
import torch

# Load the model
pipeline = Chronos2Pipeline.from_pretrained("amazon/chronos-2", device_map="cuda")

# Give list of time series.
inputs = [
    torch.tensor([1.0, 2.0, 3.0, 4.0], dtype=torch.float32), 
] 

# Embed
emb, _ = pipeline.embed(inputs)

I notice that predict_df also takes timestamp_column as input when making predictions, but the embed function doesn't. Is there a way to parse timestamps or is that not needed?

Thank you!

[vstenby]

Sorry for the bug label, couldn't add it with Q&A for some reason 😐

[abdulfatir]

No problem. Please use discussions for Q/A.

[vstenby]

Sure, will do - thanks!

[abdulfatir]

@vstenby Chronos models actually do not use timestamps internally. The predict_df expects them to correctly format sequences as time series and to prevent incorrect use of the model (e.g., using on non-uniformly spaced time series). You should compare the embed method with predict or predict_quantiles which also does not take timestamps.

[vstenby]

Amazing, thanks @abdulfatir - that makes complete sense. Thanks for the quick reply!