SKLearn model inference 'Please provide a model_fn implementation.' errors

[tom-whitehead]

I am trying to deploy a model from Sagemaker script mode, however every time I try to create an endpoint the logs include mysterious errors saying ClientError: Please provide a model_fn implementation. I have implemented model_fn as you will see below.

The notebook looks like this:

from sagemaker.sklearn.estimator import SKLearnModel
from sagemaker import get_execution_role

role = get_execution_role()

estimator = SKLearnModel(
    entry_point='inference.py',
    source_dir='/home/ec2-user/SageMaker/my_path/src/',
    model_data='s3://my_bucket/my_key/output/model.tar.gz',
    role=role,
    py_version='py3'
)

predictor = estimator.deploy(
    instance_type='ml.m4.4xlarge',
    initial_instance_count=1,
    endpoint_name='my_endpoint'
)

And my inference.py looks like this:

import json

from pandas import DataFrame

from modelling.models import MyModel
from modelling.vectorisers import MyVectoriser
from util.config import TRAINING_VARIABLE_HEADERS


def model_fn(model_dir):
    return MyModel(inference_model_dir=model_dir)


def input_fn(request_body, request_content_type):
    input_dict = json.loads(request_body)
    print(f'Processing request {input_dict}')

    raw_input = DataFrame.from_dict(input_dict['media_items_data'])

    vectoriser = MyVectoriser()
    vectorised_input = vectoriser.transform(raw_input, TRAINING_VARIABLE_HEADERS)

    return {'inputs': vectorised_input}


def predict_fn(input_dict, model):
    prediction = model.predict(input_dict['inputs'])
    return prediction.tolist()


def output_fn(prediction, accept):
    return json.dumps(prediction)

The code folder structure is like this:

src
 L modelling 
   L models.py
   L vectorisers.py
 L utils
   L config.py
 inference.py
notebooks
 L train_deploy.ipynb

I've been though the documentation and can't figure out what I am doing wrong in this case. If anyone could help I would be very grateful.

I've tried playing around with the folder architecture, but to no avail.