mlsteam-model-sdk

SDK for accessing MLSteam models

Setup

pip3 install mlsteam-model-sdk

To process encrypted model versions, install the Themis development package according to the official instrunctions. Debian/Ubuntu users have a handy installation method:

# for users that already have administrator privileges
mlsteam-model-cli install-themisdev

# for those that need privilege lifting
sudo mlsteam-model-cli install-themisdev

Usage

Initilize SDK

SDK needs to be initialized if you have not done so (replace the fields started with $):

mlsteam-model-cli init \
    --default_project_type=name \
    --default_project_val=$PROJECT_OWNER/$PROJECT_NAME

By default, the settings will be at $HOME/.mlsteam-model-sdk/cfg.ini.

If the program is running out of an MLSteam system, you may also need to setup api_token with this command instead, or by editing the api_token field in cfg.ini:

mlsteam-model-cli init \
    --api_token=$YOUR_API_TOKEN \
    --default_project_type=name \
    --default_project_val=$PROJECT_OWNER/$PROJECT_NAME

This checks whether SDK has been initialized:

$ mlsteam-model-cli check
[v] SDK has been initialized.
Configuration file: /home/ubuntu/.mlsteam-model-sdk/cfg.ini

Download a model version with SDK

from mlsteam_model_sdk.sdk.model import Model

sdk_model = Model()
sdk_model.download_model_version(model_name='model_name',
                                 version_name='version_name')

You will need administrator privileges to handle encrypted model versions. For this case, either run the Python program with sudo, or enter your password in a sudo prompt during program execution.

By default, the model version will be downloaded at $HOME/.mlsteam-model-sdk/models/download/.

This loads a model version and makes prediction:

mv = sdk_model.load_model_version(model_name='model_name',
                                  version_name='version_name')
outputs = mv.predict(inputs)

Import a model version with SDK

There is an alternative way (2-stage model deployment) to import a model version without connecting to an MLSteam system. Rather than downloading from MLSteam directory, it involves (1) getting the related file(s) from MLSteam and then (2) importing the model version from the file(s). This alternative way adds complexity of managing the model file(s) but is suitable for the situation where there is no network connection to the MLSteam system.

This example assumes the following files are locally available:

1. model version package (required)
2. package encryption key (required only for encrypted packages)

You will need administrator privileges to import an encrypted model version, as mentioned in the previous example.

To import a package:

from mlsteam_model_sdk.sdk.model import Model

sdk_model = Model(offline=True)
sdk_model.import_model_version(mv_package_file='path/to/mv/package/file',
                               enckey_file='path/to/enckey/file')
# for non-encrypted package
# sdk_model.import_model_version(mv_package_file='path/to/mv/package/file')

1. We set offline=True to have SDK operate in offline mode (without connecting to MLSteam).
2. By default, the model and version names to register are read from the package manifest. You may customize these settings with the model_name and version_name parameters.

Import a model version with CLI

This example assumes the following files are locally available:

1. model version package (required)
2. package encryption key (required only for encrypted packages)

You will need administrator privileges to import an encrypted model version, as mentioned in the previous example.

To import a package:

# for non-encrypted packages
mlsteam-model-cli mv import-local -f $PACKAGE_FILE_PATH

# for encrypted packages
mlsteam-model-cli mv import-local -f $PACKAGE_FILE_PATH -k $ENCKEY_FILE_PATH

By default, the model and version names to register are read from the package manifest. You may customize these settings with the --model_name and --version_name options.

If the operation is successful, you will find the imported pakage in local model registry:

mlsteam-model-cli mv list-local

   muuid     model_name       vuuid        version_name     puuid     packaged   encrypted      download_time
 ================================================================================================================
 __local__   ...          local-........   ...            __local__   1          ...            .....

Model Development

The SDK is equipped with Model Development Kit (MDK) to support the following activities:

1. Generating a skeleton of manifest file;
2. Validating a manifest file;
3. Creating a model version under development locally;
4. Loading a model version under development locally;
5. Registering a model version (🕒 future work).

More information could be found in Model Developer's Guide in SDK documentation.