Survey banner
The Dataiku Community is moving to a new home! We are temporary in read only mode: LEARN MORE

COMMAND EVALUATE CUSTOM MODEL

marirapellini
Level 3
COMMAND EVALUATE CUSTOM MODEL

Hi

I have created a custom model for a multiclass prediction tasks and I want to use the command EVALUATE

My model returns a list for the predictions and I get this error:

Could not run command EVALUATE: : <class 'Exception'> : Can't handle model output: <class 'list'>

 

which classes are supported as output for the model?
Thank you

0 Kudos
6 Replies
marirapellini
Level 3
Author

I also get this error: : Can't handle MLFlow model output: <class 'list'>

 

what could I do?

0 Kudos
AurelienL
Dataiker

Hello,

Can you please make sure that your MLflow model returns a pandas dataframe or a numpy array? If you don't have control over that, you can wrap the model you are using in a subclass of PythonModel, as showcased here: https://developer.dataiku.com/latest/tutorials/machine-learning/experiment-tracking/xgboost-pyfunc/i... 

 

0 Kudos
marirapellini
Level 3
Author

Hi

thank you

can a multiclass prediction model support a dataframe as output?

thank you

0 Kudos
AurelienL
Dataiker

Yes, for a multiclass model, you can return either a dataframe having only one column (the prediction), or a dataframe with one column per class, in which case each column must contain the probability for this class. For the latter, please make sure that your column order match the labels order you used in the "class_labels" parameter of set_core_metadata.

In both cases the number of rows of your dataframe must match the number of rows of the input dataframe ("model_input" in the dev guide example).

marirapellini
Level 3
Author

Thanks

i wanted to ask, referred to the link you posted yesterday, is there a way to be able to train a model and then associated to it different Wrappers?

thank you

 

0 Kudos
AurelienL
Dataiker

I'm not sure I understand your question, you would like to wrap your model in different wrappers? I'm curious about what your use case would be for that?

Anyway, I don't see why it would not work, you can simply write your different wrappers and make them all load and call the same underlying model.

0 Kudos