Appendix
Retrain and Redeploy Model
Perhaps you would to like to tweak and increase the detection quality of your model later on? You can rerun the Pipeline and increase the Sample Size (Number of Images used for training), epochs (Number of training iterations) and batch size.
When rerunning the pipeline, a new onnx model will be created in the S3 bucket /models folder and copied over the model-latest.onnx file. It will not automatically be deployed to the running model server!
|
To deploy the new model to the model server, we need to find out the new model filename. In the pipeline run view, click on the last task (model_upload) and look a the logs. You should see an entry similar to this:
initializing S3 client uploading model to models/model-2410151507.onnx model uploaded and available as "models/model-2410151507.onnx" uploading model to models/model-latest.onnx model uploaded and available as "models/model-latest.onnx"
In this case model-2410151507.onnx is the new model created by this pipeline run.
-
Now in your Openshift AI Project go to Models and click on the three dots next to the model
fedora-detection-service -
Change the Path value to your new model (e.g. models/model-2410151507.onnx)
-
This wil trigger a redeploy and you can now test the new model
| There is another more kube native way to update the model. You can actually change the yaml of the underlying OpenShift ressource. |
Model redeploy by editing Resource yaml
-
Go to the OpenShift Console select your project
-
In the menu on the left click on Home > Search
-
Now in the resources field enter
InferenceService -
Check InferenceService in the dropdown
-
Click on the
fedora-detection-service -
Click on YAML to edit
Look for the path entry
path: models/model-latest.onnx
and change it to your new model name
This opens up possibilities to use GitOps to configure your MLOps ressources by the way.