create_predictor or forward？ Difference & applicable scenarios？

[wangq326]

In MoiraiForecast, you have implemented both create_predictor and forward. What is the difference between these two methods, and if there are differences, what scenarios should they be used in respectively?

At the same time, in the directory src/uni2ts/model/moirai, there is a file named module.py. Additionally, there are three files named pretrain, finetune, and forecast, each defining a different class such as MoiraiPretrain. If one wishes to use a pre-trained model to obtain the model's output, what differences would there be in calling these three? It seems that the inputs in their forward methods are also different?

[gorold]

forward implements the standard pytorch forward function for forecasting.

create_predictor returns a gluonts Predictor object, which applies will apply various data transformations given a dataset, and return an iterator of predictions.

MoiraiPretrain is used specifically for the pre-training phase. The transformations applied are different, and will not return a nicely formatted prediction.

[wangq326]

Thank you for your reply. I can guess and understand their general functions from their names. I would like to confirm that, apart from the differences in various transformations of the dataset, they should be consistent in the computational process.

[gorold]

The checkpoint from MoiraiPretrain is used to load into MoiraiForecast, so in that sense, they are consistent. There are differences in processing the inputs to each object.