I'm running a training job using AWS SageMaker and i'm using a custom Estimator based on an available docker image from AWS. I wanted to get some feedback on whether my process is correct or not prior to deployment.
I'm running the training job in a docker container using 'local' in a SageMaker notebook instance and the training job runs successfully. However, after the job completes and saves the model to opt/model/models within the docker image, once the docker container exits, the model saved from training is lost. Ideally, i'd like to use the model for inference, however, I'm not sure about the best way of doing it. I have also tried the training job after pushing the image to ECR, but the same thing happens.
It is my understanding that the docker state is lost, once the image exits, as such, is it possible to persist the model that was produced in training in the image? One option I have thought about is saving the model output to an S3 bucket once the training job is complete, then pulling that model into another docker image for inference. Is this expected behaviour and the correct way of doing it?
I am fairly new to using SageMaker but i'd like to do it according to best practices. I've looked at a lot of the AWS documents and followed the tutorials but it doesn't seem to mention explicitly if this is how it should be done.
Thanks for any feedback on this.
You can refer to Rok's comment on saving a model file when you're using a custom estimator. That said, SageMaker built-in estimators save the model artifacts to S3. To make inferences using that model, you can either use a real-time inference endpoint for real time predictions, or a batch transformer to run inferences in batch mode. In both cases, you'll have to point the configuration to the container for inference and the model artifacts. the amazon-sagemaker-examples repository has examples for common frameworks, especially, the scikit-learn example has detailed explanations.
Also, make sure the model is being saved to /opt/ml/model/, not opt/model/models as mentioned in your question.
Related
I want to run batch predictions inside Google Cloud's vertex.ai using a custom trained model. I was able to find documentation to get online prediction working with a custom built docker image by setting up an endpoint, but I can't seem to find any documentation on what the Dockerfile should be for batch prediction. Specifically how does my custom code get fed the input and where does it put the output?
The documentation I've found is here, it certainly looks possible to use a custom model and when I tried it didn't complain, but eventually it did throw an error. According to the documentation no endpoint is required for running batch jobs.
Sagemaker is a great tool to train your models, and we save some money by using AWS spot instances. However, training jobs sometimes get stopped in the middle. We are using some mechanisms to continue from the latest checkpoint after a restart. See also the docs.
Still, how do you efficiently test such a mechanism? Can you trigger it yourself? Otherwise you have to wait until the spot instance actually ís restarted.
Also, are you expected to use the linked checkpoint_s3_uri argument or the model_dir for this? E.g. the TensorFlow estimator docs seem to suggest something model_dirfor checkpoints.
Since you can't manually terminate a sagemaker instance, run an Amazon SageMaker Managed Spot training for a small number of epochs, Amazon SageMaker would have backed up your checkpoint files to S3. Check that checkpoints are there. Now run a second training run, but this time provide the first jobs’ checkpoint location to checkpoint_s3_uri. Reference is here, this also answer your second question.
I am aware that it is possible to deploy custom containers for training jobs on google cloud and I have been able to get the same running using command.
gcloud ai-platform jobs submit training infer name --region some_region --master-image-uri=path/to/docker/image --config config.yaml
The training job was completed successfully and the model was successfully obtained, Now I want to use this model for inference, but the issue is a part of my code has system level dependencies, so I have to make some modification into the architecture in order to get it running all the time. This was the reason to have a custom container for the training job in the first place.
The documentation is only available for the training part and the inference part, (if possible) with custom containers has not been explored to the best of my knowledge.
The training part documentation is available on this link
My question is, is it possible to deploy custom containers for inference purposes on google cloud-ml?
This response refers to using Vertex AI Prediction, the newest platform for ML on GCP.
Suppose you wrote the model artifacts out to cloud storage from your training job.
The next step is to create the custom container and push to a registry, by following something like what is described here:
https://cloud.google.com/vertex-ai/docs/predictions/custom-container-requirements
This section describes how you pass the model artifact directory to the custom container to be used for interence:
https://cloud.google.com/vertex-ai/docs/predictions/custom-container-requirements#artifacts
You will also need to create an endpoint in order to deploy the model:
https://cloud.google.com/vertex-ai/docs/predictions/deploy-model-api#aiplatform_deploy_model_custom_trained_model_sample-gcloud
Finally, you would use gcloud ai endpoints deploy-model ... to deploy the model to the endpoint:
https://cloud.google.com/sdk/gcloud/reference/ai/endpoints/deploy-model
I have a model.pkl file which is pre-trained and all other files related to the ml model. I want it to deploy it on the aws sagemaker.
But without training, how to deploy it to the aws sagmekaer, as fit() method in aws sagemaker run the train command and push the model.tar.gz to the s3 location and when deploy method is used it uses the same s3 location to deploy the model, we don't manual create the same location in s3 as it is created by the aws model and name it given by using some timestamp. How to put out our own personalized model.tar.gz file in the s3 location and call the deploy() function by using the same s3 location.
All you need is:
to have your model in an arbitrary S3 location in a model.tar.gz archive
to have an inference script in a SageMaker-compatible docker image that is able to read your model.pkl, serve it and handle inferences.
to create an endpoint associating your artifact to your inference code
When you ask for an endpoint deployment, SageMaker will take care of downloading your model.tar.gz and uncompressing to the appropriate location in the docker image of the server, which is /opt/ml/model
Depending on the framework you use, you may use either a pre-existing docker image (available for Scikit-learn, TensorFlow, PyTorch, MXNet) or you may need to create your own.
Regarding custom image creation, see here the specification and here two examples of custom containers for R and sklearn (the sklearn one is less relevant now that there is a pre-built docker image along with a sagemaker sklearn SDK)
Regarding leveraging existing containers for Sklearn, PyTorch, MXNet, TF, check this example: Random Forest in SageMaker Sklearn container. In this example, nothing prevents you from deploying a model that was trained elsewhere. Note that with a train/deploy environment mismatch you may run in errors due to some software version difference though.
Regarding your following experience:
when deploy method is used it uses the same s3 location to deploy the
model, we don't manual create the same location in s3 as it is created
by the aws model and name it given by using some timestamp
I agree that sometimes the demos that use the SageMaker Python SDK (one of the many available SDKs for SageMaker) may be misleading, in the sense that they often leverage the fact that an Estimator that has just been trained can be deployed (Estimator.deploy(..)) in the same session, without having to instantiate the intermediary model concept that maps inference code to model artifact. This design is presumably done on behalf of code compacity, but in real life, training and deployment of a given model may well be done from different scripts running in different systems. It's perfectly possible to deploy a model with training it previously in the same session, you need to instantiate a sagemaker.model.Model object and then deploy it.
I have started exploring AWS SageMaker starting with these examples provided by AWS. I then made some modifications to this particular setup so that it uses the data from my use case for training.
Now, as I continue to work on this model and tuning, after I delete the inference endpoint once, I would like to be able to recreate the same endpoint -- even after stopping and restarting the notebook instance (so the notebook / kernel session is no longer valid) -- using the already trained model artifacts that gets uploaded to S3 under /output folder.
Now I cannot simply jump directly to this line of code:
bt_endpoint = bt_model.deploy(initial_instance_count = 1,instance_type = 'ml.m4.xlarge')
I did some searching -- including amazon's own example of hosting pre-trained models, but I am a little lost. I would appreciate any guidance, examples, or documentation that I could emulate and adapt to my case.
Your comment is correct - you can re-create an Endpoint given an existing EndpointConfiguration. This can be done via the console, the AWS CLI, or the SageMaker boto client.
https://docs.aws.amazon.com/cli/latest/reference/sagemaker/create-endpoint.html
https://boto3.amazonaws.com/v1/documentation/api/latest/reference/services/sagemaker.html#SageMaker.Client.create_endpoint