CONFIG.md

config.yaml schema

Below schema should be used when preparing a config.yaml file for models using the tool. Some parameters are optional and marked as such.

project_id: [project ID]
bucket_id: [GCS bucket ID]
region: [GCP region to train ML Pipeline Generator models in, on AI Platform]
cluster_name: [Name of GKE cluster hosting Kubeflow Pipelines]
cluster_zone: [Zone in which GKE cluster is deployed]
scale_tier: [compute specifications for training the model on AI Platform]
runtime_version: [AI Platform Training runtime version]
python_version: [Python version used in the model code for training]
package_name: [name for the source distribution to be uploaded to GCS]
machine_type_pred: [type of virtual machine that AI Platform Prediction uses for the nodes that serve predictions, defaults to mls1-c1-m2]

data:
	schema:
		- [schema for input & target features in the training data]
	train: [GCS location url to upload preprocessed training data]
	evaluation: [GCS location url to upload preprocessed eval data]
	prediction:
		input_data_paths:
			- [GCS location urls for prediction input data]
		input_format: [prediction input format]
		output_format: [prediction output format]

model:
	name: [unique model name, must start with a letter and only contain letters, numbers, and underscores]
	path: [local dir path to the model.py file]
	target: [target feature in training data]
	metrics: [metrics to evaluate model training on, such as “accuracy”]

model_params:
	input_args: [Any input params to be submitted with the job]
		arg_name:
			type: [data type of the arg, such as int]
			help: [short description of the arg]
			default: [default value of the arg]
	hyperparam_config: [optional; local path to hyperparam tuning config yaml. See schema here for this config file.]
	explanation: [optional; explainability features for the training job]

orchestration:
	kubeflow_url: [for KFP backend; URL of preconfigured Kubeflow instance]