The Model Format section described how to export your ML model into files. A modern ML team produces a large number of models everyday and improve them with new data or new algorithms overtime. It is critical for teams to keep track of the models they have created, make sure the production system have secure access to those models, and allow teams to share and re-use models.
Atalaya provides easy Model Management to help ML teams to achieve this goal. Make sure to Install the Atalaya CLI before running the code samples in this guide.
To upload a model in model repo "example-model", run the following command:
atalaya push example-model PATH_TO_MODEL_FILES Uploading model artifacts... Successfully pushed model example-model:2018-7-17-1531862064510-w7c3do95
You can add a specific tag for your model:
atalaya push example-model:1.0.0 PATH_TO_MODEL_FILES Uploading model artifacts... Successfully pushed model example-model:1.0.0
atalaya push command, the first parameter is model repository
name and a tag(optional). A random UUID will be generated for this model when
tag is not specififed. If the target model repository does not exist, a new
one will be created as well. See full command spec below.
Models Are Immutabile
All model artifacts are immutable once pushed to Atalaya. Every model has its own unique tag that can be referenced in the rest of the system.
Althought the tag and model artifacts can not be modified, users can always add or edit labels on their models.
Models can be removed from Web UI but not recommended. You must make sure there is not any production workload relying on the models you are deleting, and taking into consideration that you may need some of those models when rolling back to older version of your deployments.
Push Command Usage
USAGE atalaya push REPO:TAG FILE_PATH --type=MODEL_TYPE ARGUMENTS NAME Model repo name and model tag FILEPATH Model artifact path OPTIONS -d, --description=description Description for the model -l, --labels=labels Labels are key/value pairs that attached to object such as model. Example: release=test -t, --type=Tensorflow (required) [default: Tensorflow] Model type -s, --model-store=model-store The name of model store that the model is uploading to. EXAMPLES atalaya push fraud-predic:v1.2.0 ./model.pb --type=Tensorflow --labels=algo=hmm,type=creditcard atalaya push shopping-anomaly-detection ./models -t Tensorflow -d this model is ready for prod
To list all models in your organization, use
atalaya get models command:
atalaya get models Name Tag Type Created At Created By Upload Status example-model 2018-7-17-1531862064510-w7c3do95 Tensorflow 9 minutes ago chaoyu Completed example-model v2 Tensorflow 9 minutes ago chaoyu Completed example-model v1 Tensorflow a day ago chaoyu Completed mnist 2018-6-23-1532385753990-hf8xs6id Tensorflow 23 days ago chaoyu Completed summarizer 2018-6-18-1531945524788-0exb1s1t Tensorflow a month ago chaoyu Completed inception 2018-6-18-1531945503996-njz15y5f Tensorflow a month ago chaoyu Completed inception 2018-6-18-1531945275213-pdhlblg1 Tensorflow a month ago chaoyu Completed mnist 2018-7-18-1531904290125-n97k31y7 Tensorflow a month ago chaoyu Completed ...
To list models in specified model repositery:
atalaya get models example-model Name Tag Type Created At Created By Upload Status example-model 2018-7-17-1531862064510-w7c3do95 Tensorflow 9 minutes ago chaoyu Completed example-model v2 Tensorflow 9 minutes ago chaoyu Completed example-model v1 Tensorflow a day ago chaoyu Completed
To list models with specified labels and filter:
atalaya get models -l release=v1,app=test -f hotdog
Custom Model Storage
Atalaya support custom model storages for storing model artifacts. Enterprise customers can specify their own block storage target from Google Cloud Storage, AWS S3, or Azure Blob Storage as storage backend for their model repositery.
To push a model to the customized model storage, use
atalaya push example-model --model-store=my-model-store-id