ModuleNotFoundError: No module named 'azureml.training'

Question

i have this problem:

I have trained and registered a ml model via automated machine learning on the Azure Machine Learning studio platform

The result is an XGBoost model that i register without problem or waring in the workspace.

I need to deploy it to a batch endpoint, but with a custom scoring script:

def init():
global model

# AZUREML_MODEL_DIR is an environment variable created during deployment
# The path "model" is the name of the registered model's folder

model_path = os.environ["AZUREML_MODEL_DIR"]
model_file = glob.glob(f"{model_path}/*/*.pkl")[-1]

# load the model
with open(model_file, "rb") as file:
    model = pickle.load(file)

def run(mini_batch: List[str]) -> Union[List[Any], pd.DataFrame]:
results = []

probs = []
for file in mini_batch:
    rows = pd.read_csv(file)
    probs = [el[1] for el in model.predict_proba(rows)]

return pd.DataFrame(probs)

I tried these ways to load the model:

By mlflow
By pickle

The problem is always the same:

ModuleNotFoundError: No module named 'azureml.training' i tried to read the pkl file and the 'azureml.training' module is required by the model in the first row: [b'\x80\x04\x958\x05\x00\x00\x00\x00\x00\x00\x8c<azureml.training.tabular.models.forecasting_pipeline_wrapper\x94\x8c\x12RegressionPipeline\x94\x93\x94)\x81\x94}\x94 ...

I tried to install all the libraries suggested by documentation:

azureml-core
azureml-training-core
azureml-sdk[train]

Some hint?

How you are registering the model? Please give more details on that. Meanwhile try using `mlflow.pyfunc.load_model("")` — JayashankarGS, Jun 16 '23 at 14:19
I tried with mlflow.pyfunc.load_model, but same problem. The model has been found by autmo ml, it is an xgboost. To register it I used the "register" option, with mlflow type. — Alberto Bertoncini, Jun 17 '23 at 21:14