Preparing to upgrade Watson Machine Learning
If you previously installed Watson Machine Learning Accelerator, make sure to read the following before you upgrade Watson Machine Learning.
The Watson Machine
Learning Accelerator service is no longer available. After you upgrade, some
features continue to be supported.
| Feature | |
|---|---|
| Creating models using the WML API or the Deep learning experiment builder | Supported |
| Deep learning training using Deep learning experiment builder | Supported |
| Deep learning training (single, distributed, and elastic distributed training) | Supported |
| Hyperparameter optimization (HPO) and tuning (hyperparameter search) | Supported |
| Built-in hyperparameter search plug-ins | Supported |
Native Watson Machine
Learning Accelerator command line dlicm and
dlim |
Not supported |
| Watson Machine Learning Accelerator elastic distributed inference (EDI) | Not supported |
| Backup and restore of Watson Machine Learning Accelerator | Not supported |
| Importing and exporting Watson Machine Learning Accelerator data sets to Cloud Pak for Data 5.1 | Not supported |
Note:
- Features are only supported on x86_64 hardware.
- Hyperparameter search is not available in the Deep learning experimenter builder. Use the Watson Machine Learning API for hyperparameter search.
- The following built-in hyperparameter search algorithms are supported: Random, Hyperband, Tree-structured Parzen Estimator (TPE). ExperimentGridSearch is not supported.
- If you were previously using Watson Machine Learning Accelerator 4.8, you will need to recreate and retrain your models to ensure that they are using supported deep learning frameworks. For more information, see Supported software specifications.
- If you previously trained a model using Watson Machine Learning Accelerator 5.0.x, you can deploy the trained model in Watson Machine Learning for scoring. For more information on scoring, see Deploying making learning assets.
Before you upgrade
If you previously installed the Watson Machine Learning Accelerator service, you will need to uninstall Watson Machine Learning Accelerator before you upgrade. Before uninstalling Watson Machine Learning Accelerator, ensure that you back up any model definitions that you plan on recreating in Watson Machine Learning.
- Back up any model definitions and user data
- Before uninstalling Watson Machine Learning Accelerator, ensure that you back up any model definitions and user data that you plan on recreating in Watson Machine Learning. To determine the back up location of your model definitions or obtain your trained model, see Determine the mount point for deep learning experiments. Alternatively, you can download the training results using the Watson Machine Learning Accelerator API.
- Uninstall Watson Machine Learning Accelerator
-
To uninstall a previous version of Watson Machine Learning Accelerator, see:
After you upgrade
Here are some additional considerations for after you upgrade.
Hyperparameter tuning
Previously, if you used the Watson Machine Learning Accelerator API for hyperparameter tuning. You will now need to use the Watson Machine Learning API.
You can use the Watson Machine
Learning API to create a hyperparameter experiment. Make sure to
create the experiment model definition using the
hyper_parameters_optimization
section under training_references:training_references=[
{
"model_definition": {
"command": model_cmd,
"hardware_spec": hardware_spec,
"id": model_definition_id,
"parameters": parameters,
"software_spec": software_spec
},
"hyper_parameters_optimization": hyper_parameters_optimization
}
]
Example using a Random search
algorithm:
"hyper_parameters_optimization" : {
"hyper_parameters" : {
"learning_rate" : {
"max_value" : 0.1,
"min_value" : 0,
"step" : 0.002
}
},
"method" : {
"name" : "random",
"parameters" : {
"max_parallel_job_num" : 1,
"max_run_time" : -1,
"maximize_or_minimize" : "minimize",
"method_init" : {
"random_seed" : 15
},
"num_optimizer_steps" : 2,
"objective" : "loss"
}
}
},Example using a Hyperband search
algorithm:
"hyper_parameters_optimization" : {
"hyper_parameters" : {
"learning_rate" : {
"max_value" : 0.1,
"min_value" : 0,
"step" : 0.002
}
},
"method" : {
"name" : "hyperband",
"parameters" : {
"max_parallel_job_num" : 1,
"max_run_time" : -1,
"maximize_or_minimize" : "minimize",
"method_init" : {
"fidelity" : {
"name" : "epoch",
"value" : 10
},
"random_seed" : 15
},
"num_optimizer_steps" : 2,
"objective" : "loss"
}
}
},Example using a Tree-structured Parzen Estimator search
algorithm:
"hyper_parameters_optimization" : {
"hyper_parameters" : {
"learning_rate" : {
"max_value" : 0.1,
"min_value" : 0,
"step" : 0.002
}
},
"method" : {
"name" : "tpe",
"parameters" : {
"max_parallel_job_num" : 1,
"max_run_time" : -1,
"maximize_or_minimize" : "minimize",
"method_init" : {
"random_seed" : 15
},
"num_optimizer_steps" : 2,
"objective" : "loss"
}
}
},