Feature Request: Possible to Retrieve Raw Prediction Results?
JCLGeorgia opened this issue · 1 comments
Hello -
I hope all is well. Is it possible to produce raw predicted values from the model? I see the options to generate prediction plots with wandb, but is there any possibility of producing the raw data for each feature in a dataset.
The goal I have is that there is research that does not normalize their input data, so I would like to inverse transform to produce non-normalized distance metrics of the predicted results. Inputting non-normalized values to the model I believe would cause issues.
Any thoughts you have would be appreciated and thank you for the wonderful source code + information.
Best!
There are some pytorch_lightning based ways to do this but honestly I think the least confusing way is to use the trained model as a regular pytorch module. I added a predict
function to Forecaster
that handles making the unknown y_t
(target sequence variable values) and dealing with the distributional output of spacetimeformer. It takes a batch of x_c, y_c, x_t
tensors (with y_c un-normalized), then returns un-normalized predictions of y_t.