saved_path not defined in Jupyter notebook on Windows

[yashika51]

Hi,
As far as it seems it is a problem due to slashes (/) on windows. I am starting a discussion here so that other windows users can also refer it in the future.

While loading the saved model in jupyter notebooks(vs code) an error stating saved_path not defined occurs.

As a workaround when I print the complete path and pass that to load the model another error config file yml not found occurs.

I checked that the file exists but the function is not able to track that.
However, when I run the same code as a python script instead of notebook, I get the desired result. Something similar happens with another person from my pod.

What workaround can be there for using bentoml's load function+jupyter notebooks on Windows?

[parano]

Hi @yashika51 - could you share your notebook? there are many windows users using bentoML and I haven't seen this issue came up before. Based on the screenshot, it looks like you used %%writefile to create the main.py file instead of running that piece of code.

[yashika51]

Thanks for the reply @parano. I referred to examples from bentoml/gallery which were using %writefile method.
Here is the notebook:
https://nbviewer.jupyter.org/gist/yashika51/7635e3120a92869658611d410f33e83e

[parano]

I see - so the reason you are seeing this error, is because when you do !python main.py, the code is being executed in a new python session instead of the current session. That's why you can not access the saved_path variable after running !python main.py.

The first cell should use %%writefile because BentoML requires putting the main BentoService class definition to its own python file. Defining BentoService class in the main module(running in a python interactive shell or notebook) makes it hard to do code packaging:

%%writefile iris_classifier.py
#writing script for BentoML Service Class
from bentoml import env,artifacts,api, BentoService
from bentoml.artifact import SklearnModelArtifact
from bentoml.adapters import DataframeInput

@env(auto_pip_dependencies=True,
     conda_dependencies=['bentoml'])
@artifacts([SklearnModelArtifact("model")])
class IrisClassifier(BentoService):
    @api(input=DataframeInput())
  
    def predict(self,df):
        return self.artifact.model.predict(df)

However you don't need the %%writefile for the second cell, you can simply change the second cell to:

    from sklearn.svm import SVC
    from sklearn.datasets import load_iris
    from iris_classifier import IrisClassifier 

    #load data
    iris=load_iris()
    X,y=iris.data, iris.target

    #train model
    clf=SVC()
    clf.fit(X,y)

    #create an iris classifier service instance 
    service=IrisClassifier()

    #pack the newly trained service classifier
    service.pack("model",clf)

    # Save the prediction service to disk for model serving
    saved_path=service.save()

That will execute all code in current notebook's python session and give you access to the saved_path variable in following cells.

[yashika51]

Oh, Understood. Thanks for explaining that. It works fine now 😄. Should we do this small fix in the code examples? That is removing %%writefile from the second cell?

https://github.com/bentoml/gallery/blob/master/scikit-learn/iris-classifier/iris-classifier.ipynb

[parano]

Yea if you could help with a quick fix that'd be great. That example uses the !bentoml serve IrisClassifier:latest CLI command to load the service instead of using the saved_path variable so it runs without problem. But I agree it is unnecessary to write that code in a main.py file, it can be very confusing to new users.

[yashika51]

I agree. I'll make a PR for this small fix. :)