-
Notifications
You must be signed in to change notification settings - Fork 0
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
get_input_dim is unable to calculate a model's input dimension if the the model config's data_input.predictor_columns is empty #316
Comments
Ok, that makes sense. A couple of questions...
In the meantime, it sounds like the fix would be to go through and copy predictor column names from |
So I added the following code block to the SkySpark function that handles prediction and it works!
I guess that answers the short-term fix. I'll get this fix into the nrelWattileExt codebase tomorrow sometime. (Right now it is running in a local version of the function on the SkySpark server. By design, the local copy overrides what is included in nrelWattileExt.) I would still like to discuss the long-term fix of how to "lock in" the predictor columns after initial training. We can do it via this thread or at our next meeting. |
|
To frame this differently, I think the in longer term we talked about...
How exactly this happens I don't have strong feelings except to say it should not require the end user to have to modify the configuration settings themselves to make everything line up. |
The re-emergence of this bug has kicked off a new round of work, and here is the current state of things after a productive debugging session with @haneslinger. I think this ticket is more of a conversation starter and we'll implement whatever we agree upon
Background
get_input_dim uses the given config's data_input.predictor_columns as a starting point for its calculations. Therefore, loading a pre-trained model whose config contains empty predictor_columns (or, more specifically, an empty array) leads to mismatched shapes from what the underlying torch model expects and what it receives.
The recommended short-term fix
We should ensure that the configurations of pre-trained models contain predictor columns.
The long-term fix
The team should discuss, and we'll settle on an approach that works for everyone.
The text was updated successfully, but these errors were encountered: