Cannot start Model Error Analysis webapp

ihb

Hi Dataiku Community,

I tried installing the plugin for Model Error Analysis and I'm having some issues (happening with other plugins as well: Model Stress Test, Model Drift Monitoring, Generalized Linear Models). I am getting a backend error when starting the webapp form the model view.

Traceback (most recent call last):
  File "<string>", line 32, in get_original_model_info
  File "/opt/dataiku/python/dataiku/doctor/posttraining/model_information_handler.py", line 158, in from_full_model_id
    return PredictionModelInformationHandler._for_trained_model(*match.groups())
  File "/opt/dataiku/python/dataiku/doctor/posttraining/model_information_handler.py", line 172, in _for_trained_model
    core_params = dkujson.load_from_filepath(osp.join(session_folder, "core_params.json"))
  File "/opt/dataiku/python/dataiku/core/dkujson.py", line 36, in load_from_filepath
    with open(filepath) as fp:
FileNotFoundError: [Errno 2] No such file or directory: '/opt/dataiku/analysis-data/MYPROJECT1/uSs2XOTr/bXx5Fw8j/sessions/s8/core_params.json'

The error seems to occur when calling PredictionModelInformationHandler.from_full_model_id() from the package dataiku.doctor.posttraining.model_information_handler

First, I assumed the reason what that the params file was not transferred to the pod because it's not in the pod. But, after trying to run the code section from the plugin locally, I am getting the same error even though file is available on that path.

Has anyone faced this issue?

AgatheG

Hi ihb,

Apologies, I got a bit mixed up in my original answer... The issue seems to actually stem from the model view running in a container (not that the model was trained in a container).

In order to prevent model views from being run in a container, the administrator of your DSS instance should change the following parameter (in the instance's Administration page > Settings top tab > Containerized execution left tab > Virtual webapp backends section):

Note that you have to set Behavior to Run in container - providing None for Container will guarantee local execution.

Hope this helps!

Agathe

AgatheG

Hi ihb,

Did you perchance train your model with a containerized execution (docker, K8S)? Model views do not support containerized execution for now.

ihb

Hi AgatheG,

That is indeed the case - the models were trained on Kubernetes. Is there no way to use model views locally after the model was trained in containerized execution? These models are quite large and it's not ideal to train them on the host.