Model metadata and validation schema¶
model-metadata.yaml file is used to specify additional information about a custom task or a custom inference model, such as:
Supported input/output data types that validate, when composing a blueprint, whether a task's input/output requirements match the neighboring tasks.
The environment ID/model ID of a task or model when running
To define metadata, create a
model-metadata.yaml file and put it in the top level of the task/model directory. In most cases it can be skipped, but it is required for custom transform tasks when a custom task outputs non-numeric data.
You can see a full example. Note that
model-metadata.yaml is located in the same folder as
The sections below show how to define metadata for custom models and tasks.
General metadata parameters¶
The following table describes options that are available to tasks and/or inference models. The parameters are required when using
drum push to supply information about the model/task/version to create. Some of the parameters are also required outside of
drum push, for compatibility reasons.
modelID parameter adds a new version to a pre-existing custom model or task with the specified ID. Because of this, all options that configure a new base-level custom model or task are ignored when passed alongside this parameter. However, at this time these parameters still must be included.
|Option||When required||Task or inference model||Description|
||Always||Both||A string, preferably unique for easy searching, that
||Always||Both||A string, either
||Always||Both||A hash of the execution environment to use while running your custom model or task. You can find a list of available execution environments in Model Registry > Custom Model Workshop > Environments. Expand the environment and click on the Environment Info tab to view and copy the file ID. Required for
||Always||Both||A string indicating the type of target. Must be one of:
||Optional||Both||After creating a model or task, it is best practice to use versioning to add code while iterating. To create a new version instead of a new model or task, use this field to link the custom model/task you created. The ID (hash) is available from the UI, via the URL of the custom model or task. Used with
||Optional||Both||A searchable field. If
||Optional||Both||Specifies whether the model version you are creating should be a major (
||Always||Model||A string indicating the column in your data that the model is predicting.|
||For binary classification models||Model||When your model predicts probability, the
||Optional (binary classification models only).||Model||The cutoff point between 0 and 1 that dictates which label will be chosen as the predicted label.|
||Optional||Task||A hash with the ID of the project (PID) to train the model or version on. When using
Validation schema and fields¶
The schema validation system, which is defined under the
typeSchema field in
model_metadata.yaml, is used to define the expected input and output data requirements for a given custom task. By including the optional
output_requirements fields, you can specify exactly the kind of data a custom task expects or outputs. DataRobot displays the specified conditions in the blueprint editor to indicate whether the neighboring tasks match. It also uses them during blueprint training to validate whether the task's data format matches the conditions. Supported conditions include:
- data type
- data sparsity
- number of columns
- support of missing values
Be aware that
output_requirements are only supported for custom transform tasks and must be omitted for estimators.
The sections below describe allowed conditions and values. Unless noted otherwise, a single entry is all that is required for input and/or output requirements.
data_types field specifies the data types that are expected, or those that are specifically disallowed. A single data type or a list is allowed for
input_requirements; only a single data type is allowed as
Allowed values are NUM, TXT, IMG, DATE, CAT, DATE_DURATION, COUNT_DICT, and GEO.
The conditions used for
- EQUALS: All of the listed data types are required in the dataframe. Missing or unexpected types raise an error.
- IN: All of the listed data types are supported, but not all are required to be present.
- NOT_EQUALS: The data type for the input dataframe may not be this value.
- NOT_IN: None of the listed data types are supported by the task.
sparse field defines whether the task supports sparse data as an input or if the task can create output data that is in a sparse format.
- A condition of EQUALS must always be included in sparsity specifications.
For input, the following values apply:
- FORBIDDEN: The task cannot handle a sparse matrix format, and will fail if one is provided.
- SUPPORTED: The model must support both a dense dataframe and a sparse dataframe in CSR format. Either could be passed in from preceding tasks.
- REQUIRED: This task only supports a sparse matrix as input and cannot use a dense matrix. DRUM will load the matrix into a sparse dataframe.
For task output, the following values apply:
- NEVER: The task can never output a sparse dataframe.
- DYNAMIC: The task can output either a dense or sparse matrix.
- ALWAYS: The task will always output a sparse matrix.
- IDENTITY: The task can output either a sparse or dense matrix, and the sparsity will match the input matrix.
number_of_columns field specifies whether a specific minimum or maximum number of columns is required. The value should be a non-negative integer.
For time-consuming tasks, specifying a maximum number of columns can help keep performance reasonable. The
number_of_columns field allows multiple entries to create ranges of allowed values. Some conditions only allow a single entry (see the example).
The conditions used for
number_of_columns in a dataframe are:
- EQUALS: The number of columns must exactly match the value. No additional conditions allowed.
- IN: Multiple possible acceptable values are possible. The values are provided as a list in the value field. No additional conditions allowed.
- NOT_EQUALS: The number of columns must not be the specified value.
- GREATER_THAN: The number of columns must be greater than the value provided.
- LESS_THAN: The number of columns must be less than the value provided.
- NOT_GREATER_THAN: The number of columns must be less than or equal to the value provided.
- NOT_LESS_THAN: The number of columns must be greater than or equal to the value provided.
The value must be a non-negative integer.
contains_missing field specifies whether a task can accept missing data or whether a task can output missing values.
- A condition of EQUALS must always be used.
For input, the following values apply to the input dataframe:
- FORBIDDEN: The task cannot accept missing values/NA.
- SUPPORTED: The task is capable of dealing with missing values.
For task output, the following values apply:
- NEVER: The task can never output missing values.
- DYNAMIC: The task can output missing values.
name: default-transform-model-metadata type: training targetType: transform typeSchema: input_requirements: - field: data_types condition: IN value: - NUM - CAT - TXT - DATE - DATE_DURATION - field: sparse condition: EQUALS value: SUPPORTED - field: contains_missing condition: EQUALS value: SUPPORTED output_requirements: - field: data_types condition: EQUALS value: NUM - field: sparse condition: EQUALS value: DYNAMIC - field: contains_missing condition: EQUALS value: DYNAMIC
The default output data type is NUM. If any of these values are not appropriate for the task, a schema must be supplied in
model-metadata.yaml (which is required for custom transform tasks that output non-numeric data).
Running checks locally¶
drum fit or
drum push, the full set of validation is run automatically. Verification first checks that the supplied
typeSchema items meet the required format. Any format issues must be addressed before the task can be trained locally or on DataRobot. After format validation, the input dataset used for fit is compared against the supplied
input_requirements specifications. Following task training, the output of the task is compared to the
output_requirements and an error is reported if a mismatch is present.
During task development, it might be useful to disable validation. To ignore errors, use the following with
drum fit or