Support SELECT expr; which does not require FROM table

[xzdandy]

Address the SELECT Forecast(12) FROM AirData; to SELECT Forecast(12); in #1081

• update parser, binder, optimizer, and executor to allow project without children.
• update forecasting test cases and documentation.
• add unit test and short integration test for SELECT expr;.
• add documentation that we support SELECT expr;.

[xzdandy]

Discussed offline. We will send a dummy batch instead of checking batch is None everywhere.

[xzdandy]

@gaurav274 @jiashenC Please review. We should be ready for a forecasting notebook after this PR.

[jiashenC]

Thanks, the PR looks good to me.

Just curious about the workflow of the forecasting model for my own knowledge.

• During training (i.e., CREATE FUNCTION), the forecasting model reads a table and memorizes all of the data from the table to its internal state (e.g., weights).
• During inference, the forecasting model outputs the prediction from its internal state.

Is this the reason why the SeqScan is no longer needed during inference?

[xzdandy]

Thanks, the PR looks good to me.

Just curious about the workflow of the forecasting model for my own knowledge.

• During training (i.e., CREATE FUNCTION), the forecasting model reads a table and memorizes all of the data from the table to its internal state (e.g., weights).
• During inference, the forecasting model outputs the prediction from its internal state.

Is this the reason why the SeqScan is no longer needed during inference?

Yes. It always works on the original dataset or table.

[jiashenC]

Thanks, the PR looks good to me.
Just curious about the workflow of the forecasting model for my own knowledge.

• During training (i.e., CREATE FUNCTION), the forecasting model reads a table and memorizes all of the data from the table to its internal state (e.g., weights).
• During inference, the forecasting model outputs the prediction from its internal state.

Is this the reason why the SeqScan is no longer needed during inference?

Yes. It always works on the original dataset or table.

I see. Then, if the user wants the same model to work with a different dataset, the user needs to call CREATE FUNCTION again. Is that correct?

[xzdandy]

Thanks, the PR looks good to me.
Just curious about the workflow of the forecasting model for my own knowledge.

• During training (i.e., CREATE FUNCTION), the forecasting model reads a table and memorizes all of the data from the table to its internal state (e.g., weights).
• During inference, the forecasting model outputs the prediction from its internal state.

Is this the reason why the SeqScan is no longer needed during inference?

Yes. It always works on the original dataset or table.

I see. Then, if the user wants the same model to work with a different dataset, the user needs to call CREATE FUNCTION again. Is that correct?

That is correct. On the other hand, If I understand correctly, for forecasting, the same model can not work for other datesets. It is trained for a specific dataset. @americast who should have more knowledge in this space.

[americast]

Thanks @xzdandy for the PR aned @jiashenC for the review. Right, a function would be linked to a trained model linked to a specific dataset. Thus, while calling the function, we would not need to pass any dataset information.