zeroshot_trainer.predict and TimeSeriesForecastingPipeline

[NING-CSU]

I have some concerns regarding inference in a zero-shot scenario using TTM.

In the ttm_getting_started, predictions can be obtained using zeroshot_trainer.predict, but when I load a custom dataset, the output data is normalized by default. How can I obtain the denormalized data? Is there a scaler available? Additionally, I noticed that predictions.label_ids is empty; how can I get the label data (ground truth)?

Furthermore, it seems that inference can also be performed using TimeSeriesForecastingPipeline. What are the differences between TimeSeriesForecastingPipeline and zeroshot_trainer.predict? When using TimeSeriesForecastingPipeline for inference, does the input data need to be normalized?

These questions are confusing to me. I look forward to your response.

[wgifford]

Hi @NING-CSU trainer.predict works with torch datasets and outputs tensors. For higher-level functionality, we created the TimeSeriesForecastingPipeline. The pipeline takes dataframes as input and outputs dataframes. If initialized with a preprocessor, it can also perform the inverse normalization. An example can be found here: https://github.com/ibm-granite-community/granite-timeseries-cookbook/blob/main/recipes/Time_Series/Preprocessor_Use_and_Performance_Evaluation.ipynb