如何在本地环境中使用来自 Amazon Sagemaker 的 a.joblib model？

Question

I created a model in AWS using Sagemaker. I downloaded model.joblib to my machine. I am trying to use it to make some predictions. I can load the file:

import joblib
import mlio
import sklearn
filename=r"C:\Users\benki\Downloads\model.tar\model.joblib"
loaded_model = joblib.load(filename)

However, I am not sure where to go from here. I've tried to score and predict 'loaded_model' but I only get error messages explaining that 'loaded_model' does not have these attributes. 'loaded_model' is the type: sagemaker_sklearn_extension.externals.automl_transformer.AutoMLTransformer

In AWS from a Sagemaker Jupyter Notebook instance, I can make predictions with the following:

endpoint_name = "My_Model"  
#print(f"Note: Invoking Endpoint: {endpoint_name}")
content_type = "text/csv"                                        
accept = "text/csv"  

# create for loop to create results, one at a time
predict=[]
for sample in payloads:
    response = client.invoke_endpoint(
    EndpointName=endpoint_name,
    ContentType=content_type,
    Body=sample
    )
    #print('inference complete')
    inference = (response['Body'].read().decode('ascii'))
    predict.append((sample,inference))

How do I engage this joblib model?

Answer 1

Taking a look at both the exact error log and your entire local inference code might prove helpful in this case. However, I do have some ideas for you to check in the meantime:

• SageMaker endpoints are deployed over full application enviroments, where code and artifacts are integrated. It is likely that you lack some of the resources SM uses to set up inference, such as transformer definitions or pre-processing steps.

• Here you can see the source code of the AutoMLTransformer resource. It is worth noting that its constructor requires access to feature and target Sklearn transformer objects containing fit() and transform() methods. Are these part of your local pipeline?

• After training a model in SageMaker, some artifacts are created and stored in a user-defined key path in Amazon S3 (tipically as s3://bucket/your_artifacts_path/model.tar.gz ). When decompressing, you should be able to see a model/ folder containing the model.joblib artifact and the code/ sub-folder. Inside the latter, automatically generated reference Python scripts for both data pre-processing and serving are available. Sure, some parts of them might not run as-is in your local environment, but try using the serving script as a reference to determine how to pre-process incoming records and make calls against the model object.

Let me know if you find something else!