A first attempt at simplifying the code for models has been to code a new parent BaseTSFMSolver class in which most of the logics for forecasting is located (such that each new model would just need to inherit + implement batch_forecast and a model loader).
The next step would be to do the same for embed and time_embed, see discussion #31
That means having batch_embed and batch_time_embed (signature defined in BaseTSFMSolver with tensor shapes specified in the docs, actual implementation deferred per model).
Once this is done, a full pass on the whole set of available models should be done to make sure they stick (if it makes sense at least) to that behaviour.
Also, Chronos vs Chronos-v2 is a mess at the time, this will need to be improved at some point.
A first attempt at simplifying the code for models has been to code a new parent
BaseTSFMSolverclass in which most of the logics for forecasting is located (such that each new model would just need to inherit + implementbatch_forecastand a model loader).The next step would be to do the same for
embedandtime_embed, see discussion #31That means having
batch_embedandbatch_time_embed(signature defined inBaseTSFMSolverwith tensor shapes specified in the docs, actual implementation deferred per model).Once this is done, a full pass on the whole set of available models should be done to make sure they stick (if it makes sense at least) to that behaviour.
Also, Chronos vs Chronos-v2 is a mess at the time, this will need to be improved at some point.