Documentation & Fine tuning

[duvi86]

Hello,

Great contribution! Where would be defined what all parameters represent and how to choose which one is the best for us? Is there a way to enter a multiple time series forecasting way? Typically, we have 1000 timeseries, somehow correlated or not. IS there an aggregate way to do this more effectively than 1 timeseries at a time?
How can we fine tune the model to our data?

Thanks!

[lostella]

@duvi86 The docstring of the predict method contains a description of the parameters. You can leave the default ones to start, and only provide the time series and required prediction_length. A batch of multiple time series can be passed in in a list: note that the maximum allowed batch size depends on the amount of available memory (GPU or CPU) and prediction length and number of samples; this will typically be faster than predicting one series at a time. Note that each series will be predicted independently of the others (the model is univariate).

For fine tuning, please see #12.

[rasyousefi]

Can we finetune the model with our own dataset ?

[lostella]

For fine tuning, please see #12

[rasyousefi]

Can we add missing samples into the input time series provided to the model ? If yes, how can we do that ?
Does the preprocessing tools looks for input time series at regular intervals ?

[lostella]

Yes, just use torch.nan in the series, wherever observations are missing. The model assumes that input time series have observations at regular intervals.

[abdulfatir]

We have released the training/fine-tuning code. Check out the examples here: https://github.com/amazon-science/chronos-forecasting/tree/main/scripts