* Revision of getting started guide up to Batch scoring. Also new diagam and fix to ARM template to remove region restrictions.
* Detail on Batch scoring for Getting Started and additional debug message in the copy to ease of diagnosing issues
* Tweaked text and added a NOQA for message
* Clarified/Fixed getting started instructions for WebApp/AppService deployment
Co-authored-by: Joao Pedro Martins <joaopedro.martins@microsoft.com>
* Revision of getting started guide up to Batch scoring. Also new diagam and fix to ARM template to remove region restrictions.
* Detail on Batch scoring for Getting Started and additional debug message in the copy to ease of diagnosing issues
* Tweaked text and added a NOQA for message
Co-authored-by: Joao Pedro Martins <joaopedro.martins@microsoft.com>
* Change AKS deployment configuration
Deployment config changed from 1CPU/4GB to 0.5CPU/2GB so it fits a AKS created with default parameters
* Update custom_model.md
Co-authored-by: João Pedro Martins <lokijota@users.noreply.github.com>
TRAIN_SCRIPT_PATH value updated from 'training/train.py' to 'training/train_aml.py'.
This is aligned with /.pipelines/diabetes_regression-variables-template.yml.
* development_setup.md update
development_setup.md updated to use install_requirements.sh.
See #158:
> Use conda rather than pip packages when possible (as recommended in AML docs).
> Dev environment is hence also constrained to conda (no more pip install -r requirements.txt).
* Content of install_requirements.sh deleted
* build_train_pipeline.py filename fixed
* build_train_pipeline.py filename fixed
* docs
* fix pipeline status badge and tf naming uniqueness
* add a note about how to change the name of the pipeline
* extra clarification on workspace connection
Recently the step to get the model version had a name associated with the step. We need to also update the references in future step to use the step name as a prefix. This resulted in the MODEL_VERSION variable causing failures in each CD deployment step.
This did not show up in CI because the MODEL_VERSION var is hard coded in the variable group. We should also remove that.
A bug surfaced where first time evaluation of a model fails due to the
Model constructor throwing if the model does not exist.
Looking deeper, we see that most calls to get_model expect a possible
None response and check at the call site. Unfortunately we get the same
WebserviceException class for a model not being found as we do a REST
error or similar.
This change is a stopgap mitigation to restore compatibility with the
existing callers, and compromises by allowing the model version
dependent behavior to continue passing on exceptions.
In a future follow up we should settle on a convention and allow version
checks to propagate failure while still giving the possibility for
handling a service exception in the caller.
- Tied SDK version to 1.2.x as with conda_dependencies.yml
- Lock versions to point updates
- Kept the rest of the deps manually specified to keep image size small and minimize regressions
Azure deprecated their top level meta-package which led to a deprecation error. We don't actually need this top level package.
I took the opportunity to clean up the conda deps using the dependency tree.
- Trimmed the package list
- Explicitly list pip to avoid conda warning
- Use azureml-defaults for WebApp dependencies
- Lock azureml-sdk and azureml-defaults versions
- Add comments for dependencies
* Add vars template to canary pipeline
* Enable ACR authentication on AKS using a service principal
- Upgrade helm version to 3.1.1
- Remove ACR secret from the abtest-model deployment
João Pedro Martins
Tim Johnson
Katzmann1983
JennyPopova
Jenny So
Daniel Heinze
SATO Naoki (Neo)
Tom Care
jitghosh
My
David Tesar
jotaylo
Kosuke Fujimoto
Stephanie Visser
Sushant Divate
Olivier Martin
Cindy Weng
Eugene Fedorenko
Claus M