How to use a pretrained model from s3 to predict some data?

I have trained a semantic segmentation model using the sagemaker and the out has been saved to a s3 bucket. I want to load this model from the s3 to predict some images in sagemaker.

I know how to predict if I leave the notebook instance running after the training as its just an easy deploy but doesn’t really help if I want to use an older model.

I have looked at these sources and been able to come up with something myself but it doesn’t work hence me being here:

https://course.fast.ai/deployment_amzn_sagemaker.html#deploy-to-sagemaker https://aws.amazon.com/getting-started/tutorials/build-train-deploy-machine-learning-model-sagemaker/

https://sagemaker.readthedocs.io/en/stable/pipeline.html

https://github.com/awslabs/amazon-sagemaker-examples/blob/master/advanced_functionality/inference_pipeline_sparkml_xgboost_abalone/inference_pipeline_sparkml_xgboost_abalone.ipynb

My code is this:

from sagemaker.pipeline import PipelineModel
from sagemaker.model import Model

s3_model_bucket = 'bucket'
s3_model_key_prefix = 'prefix'
data = 's3://{}/{}/{}'.format(s3_model_bucket, s3_model_key_prefix, 'model.tar.gz')
models = ss_model.create_model() # ss_model is my sagemaker.estimator

model = PipelineModel(name=data, role=role, models= [models])
ss_predictor = model.deploy(initial_instance_count=1, instance_type='ml.c4.xlarge')

Answer

You can actually instantiate a Python SDK model object from existing artifacts, and deploy it to an endpoint. This allows you to deploy a model from trained artifacts, without having to retrain in the notebook. For example, for the semantic segmentation model:

trainedmodel = sagemaker.model.Model(
    model_data='s3://...model path here../model.tar.gz',
    image='685385470294.dkr.ecr.eu-west-1.amazonaws.com/semantic-segmentation:latest',  # example path for the semantic segmentation in eu-west-1
    role=role)  # your role here; could be different name

trainedmodel.deploy(initial_instance_count=1, instance_type='ml.c4.xlarge')

And similarly, you can instantiate a predictor object on a deployed endpoint from any authenticated client supporting the SDK, with the following command:

predictor = sagemaker.predictor.RealTimePredictor(
    endpoint='endpoint name here',
    content_type='image/jpeg',
    accept='image/png')

More on those abstractions:

• Model: https://sagemaker.readthedocs.io/en/stable/model.html
• Predictor: https://sagemaker.readthedocs.io/en/stable/predictors.html