Class TemporalFusionTransformerForecastingTrainingJob (1.63.0)

TemporalFusionTransformerForecastingTrainingJob(
    display_name: typing.Optional[str] = None,
    optimization_objective: typing.Optional[str] = None,
    column_specs: typing.Optional[typing.Dict[str, str]] = None,
    column_transformations: typing.Optional[
        typing.List[typing.Dict[str, typing.Dict[str, str]]]
    ] = None,
    project: typing.Optional[str] = None,
    location: typing.Optional[str] = None,
    credentials: typing.Optional[google.auth.credentials.Credentials] = None,
    labels: typing.Optional[typing.Dict[str, str]] = None,
    training_encryption_spec_key_name: typing.Optional[str] = None,
    model_encryption_spec_key_name: typing.Optional[str] = None,
)

Class to train Temporal Fusion Transformer (TFT) forecasting models.

The TemporalFusionTransformerForecastingTrainingJob class uses the Temporal Fusion Transformer (TFT) training method to train and run a forecasting model. The TFT training method implements an attention-based deep neural network (DNN) model that uses a multi-horizon forecasting task to produce predictions.

For sample code that shows you how to use `TemporalFusionTransformerForecastingTrainingJob, see the Create a training pipeline forecasting temporal fusion transformer sample on GitHub.

Properties

create_time

Time this resource was created.

display_name

Display name of this resource.

encryption_spec

Customer-managed encryption key options for this Vertex AI resource.

If this is set, then all resources created by this Vertex AI resource will be encrypted with the provided encryption key.

end_time

Optional. The time when the training job entered the PIPELINE_STATE_SUCCEEDED, PIPELINE_STATE_FAILED, or PIPELINE_STATE_CANCELLED state.

error

Optional. Detailed error information for this training job resource. Error information is created only when the state of the training job is PIPELINE_STATE_FAILED or PIPELINE_STATE_CANCELLED.

evaluated_data_items_bigquery_uri

BigQuery location of exported evaluated examples from the Training Job

Returns
Type	Description
`str`	BigQuery uri for the exported evaluated examples if the export feature is enabled for training. None: If the export feature was not enabled for training.

gca_resource

The underlying resource proto representation.

has_failed

Returns true if the training job failed, otherwise false.

labels

User-defined labels containing metadata about this resource.

Read more about labels at https://goo.gl/xmQnxf

name

Name of this resource.

resource_name

Full qualified resource name.

start_time

Optional. The time when the training job first entered the PIPELINE_STATE_RUNNING state.

state

Current training state.

update_time

Time this resource was last updated.

Methods

TemporalFusionTransformerForecastingTrainingJob

TemporalFusionTransformerForecastingTrainingJob(
    display_name: typing.Optional[str] = None,
    optimization_objective: typing.Optional[str] = None,
    column_specs: typing.Optional[typing.Dict[str, str]] = None,
    column_transformations: typing.Optional[
        typing.List[typing.Dict[str, typing.Dict[str, str]]]
    ] = None,
    project: typing.Optional[str] = None,
    location: typing.Optional[str] = None,
    credentials: typing.Optional[google.auth.credentials.Credentials] = None,
    labels: typing.Optional[typing.Dict[str, str]] = None,
    training_encryption_spec_key_name: typing.Optional[str] = None,
    model_encryption_spec_key_name: typing.Optional[str] = None,
)

Constructs a Forecasting Training Job.

Parameters
Name	Description
`display_name`	`str` Optional. The user-defined name of this TrainingPipeline.
`optimization_objective`	`str` Optional. Objective function the model is to be optimized towards. The training process creates a Model that optimizes the value of the objective function over the validation set. The supported optimization objectives: "minimize-rmse" (default) - Minimize root-mean-squared error (RMSE). "minimize-mae" - Minimize mean-absolute error (MAE). "minimize-rmsle" - Minimize root-mean-squared log error (RMSLE). "minimize-rmspe" - Minimize root-mean-squared percentage error (RMSPE). "minimize-wape-mae" - Minimize the combination of weighted absolute percentage error (WAPE) and mean-absolute-error (MAE). "minimize-quantile-loss" - Minimize the quantile loss at the defined quantiles. (Set this objective to build quantile forecasts.)
`column_specs`	`Dict[str, str]` Optional. Alternative to column_transformations where the keys of the dict are column names and their respective values are one of AutoMLTabularTrainingJob.column_data_types. When creating transformation for BigQuery Struct column, the column should be flattened using "." as the delimiter. Only columns with no child should have a transformation. If an input column has no transformations on it, such a column is ignored by the training, except for the targetColumn, which should have no transformations defined on. Only one of column_transformations or column_specs should be passed.
`column_transformations`	`List[Dict[str, Dict[str, str]]]` Optional. Transformations to apply to the input columns (i.e. columns other than the targetColumn). Each transformation may produce multiple result values from the column's value, and all are used for training. When creating transformation for BigQuery Struct column, the column should be flattened using "." as the delimiter. Only columns with no child should have a transformation. If an input column has no transformations on it, such a column is ignored by the training, except for the targetColumn, which should have no transformations defined on. Only one of column_transformations or column_specs should be passed. Consider using column_specs as column_transformations will be deprecated eventually.
`project`	`str` Optional. Project to run training in. Overrides project set in aiplatform.init.
`location`	`str` Optional. Location to run training in. Overrides location set in aiplatform.init.
`credentials`	`auth_credentials.Credentials` Optional. Custom credentials to use to run call training service. Overrides credentials set in aiplatform.init.
`labels`	`Dict[str, str]` Optional. The labels with user-defined metadata to organize TrainingPipelines. Label keys and values can be no longer than 64 characters (Unicode codepoints), can only contain lowercase letters, numeric characters, underscores and dashes. International characters are allowed. See https://goo.gl/xmQnxf for more information and examples of labels.
`training_encryption_spec_key_name`	`Optional[str]` Optional. The Cloud KMS resource identifier of the customer managed encryption key used to protect the training pipeline. Has the form: `projects/my-project/locations/my-region/keyRings/my-kr/cryptoKeys/my-key`. The key needs to be in the same region as where the compute resource is created. If set, this TrainingPipeline will be secured by this key. Note: Model trained by this TrainingPipeline is also secured by this key if `model_to_upload` is not set separately. Overrides encryption_spec_key_name set in aiplatform.init.
`model_encryption_spec_key_name`	`Optional[str]` Optional. The Cloud KMS resource identifier of the customer managed encryption key used to protect the model. Has the form: `projects/my-project/locations/my-region/keyRings/my-kr/cryptoKeys/my-key`. The key needs to be in the same region as where the compute resource is created. If set, the trained Model will be secured by this key. Overrides encryption_spec_key_name set in aiplatform.init.

Exceptions
Type	Description
`ValueError`	If both column_transformations and column_specs were provided.

cancel

cancel() -> None

Asynchronously attempts to cancel a training job.

The server makes a best effort to cancel the job, but the training job can't always be cancelled. If the training job is canceled, its state transitions to CANCELLED and it's not deleted.

Exceptions
Type	Description
`RuntimeError`	If this training job isn't running, then a runtime error is raised.

delete

delete(sync: bool = True) -> None

Deletes this Vertex AI resource. WARNING: This deletion is permanent.

Parameter
Name	Description
`sync`	`bool` Whether to execute this deletion synchronously. If False, this method will be executed in concurrent Future and any downstream object will be immediately returned and synced when the Future has completed.

done

done() -> bool

Method indicating whether a job has completed.

get

get(
    resource_name: str,
    project: typing.Optional[str] = None,
    location: typing.Optional[str] = None,
    credentials: typing.Optional[google.auth.credentials.Credentials] = None,
) -> google.cloud.aiplatform.training_jobs._TrainingJob

Gets a training job using the resource_name that's passed in.

Parameters
Name	Description
`resource_name`	`str` Required. A fully-qualified resource name or ID.
`project`	`str` Optional. The name of the Google Cloud project to retrieve the training job from. This overrides the project that was set by `aiplatform.init`.
`location`	`str` Optional. The Google Cloud region from where the training job is retrieved. This region overrides the region that was set by `aiplatform.init`.
`credentials`	`auth_credentials.Credentials` Optional. The credentials that are used to upload this model. These credentials override the credentials set by `aiplatform.init`.

Exceptions
Type	Description
`ValueError`	A `ValueError` is raised if the task definition of the retrieved training job doesn't match the custom training task definition.

get_model

get_model(sync=True) -> google.cloud.aiplatform.models.Model

Returns the Vertex AI model produced by this training job.

Parameter
Name	Description
`sync`	`bool` If set to `true`, this method runs synchronously. If `false`, this method runs asynchronously.

Exceptions
Type	Description
`RuntimeError`	A runtime error is raised if the training job failed or if a model wasn't produced by the training job.

list

list(
    filter: typing.Optional[str] = None,
    order_by: typing.Optional[str] = None,
    project: typing.Optional[str] = None,
    location: typing.Optional[str] = None,
    credentials: typing.Optional[google.auth.credentials.Credentials] = None,
) -> typing.List[google.cloud.aiplatform.base.VertexAiResourceNoun]

Lists all instances of this training job resource.

The following shows an example of how to call CustomTrainingJob.list:

aiplatform.CustomTrainingJob.list(
    filter='display_name="experiment_a27"',
    order_by='create_time desc'
)

Parameters
Name	Description
`filter`	`str` Optional. An expression for filtering the results of the request. For field names, snake_case and camelCase are supported.
`order_by`	`str` Optional. A comma-separated list of fields used to sort the returned traing job resources. The defauilt sorting order is ascending. To sort by a field name in descending order, use `desc` after the field name. The following fields are supported: `display_name`, `create_time`, `update_time`.
`project`	`str` Optional. The name of the Google Cloud project to which to retrieve the list of training job resources. This overrides the project that was set by `aiplatform.init`.
`location`	`str` Optional. The Google Cloud region from where the training job resources are retrieved. This region overrides the region that was set by `aiplatform.init`.
`credentials`	`auth_credentials.Credentials` Optional. The credentials that are used to retrieve list. These credentials override the credentials set by `aiplatform.init`.

Returns
Type	Description
`List[VertexAiResourceNoun]`	A list of training job resources.

run

run(
    dataset: google.cloud.aiplatform.datasets.time_series_dataset.TimeSeriesDataset,
    target_column: str,
    time_column: str,
    time_series_identifier_column: str,
    unavailable_at_forecast_columns: typing.List[str],
    available_at_forecast_columns: typing.List[str],
    forecast_horizon: int,
    data_granularity_unit: str,
    data_granularity_count: int,
    training_fraction_split: typing.Optional[float] = None,
    validation_fraction_split: typing.Optional[float] = None,
    test_fraction_split: typing.Optional[float] = None,
    predefined_split_column_name: typing.Optional[str] = None,
    timestamp_split_column_name: typing.Optional[str] = None,
    weight_column: typing.Optional[str] = None,
    time_series_attribute_columns: typing.Optional[typing.List[str]] = None,
    context_window: typing.Optional[int] = None,
    export_evaluated_data_items: bool = False,
    export_evaluated_data_items_bigquery_destination_uri: typing.Optional[str] = None,
    export_evaluated_data_items_override_destination: bool = False,
    quantiles: typing.Optional[typing.List[float]] = None,
    validation_options: typing.Optional[str] = None,
    budget_milli_node_hours: int = 1000,
    model_display_name: typing.Optional[str] = None,
    model_labels: typing.Optional[typing.Dict[str, str]] = None,
    model_id: typing.Optional[str] = None,
    parent_model: typing.Optional[str] = None,
    is_default_version: typing.Optional[bool] = True,
    model_version_aliases: typing.Optional[typing.Sequence[str]] = None,
    model_version_description: typing.Optional[str] = None,
    additional_experiments: typing.Optional[typing.List[str]] = None,
    hierarchy_group_columns: typing.Optional[typing.List[str]] = None,
    hierarchy_group_total_weight: typing.Optional[float] = None,
    hierarchy_temporal_total_weight: typing.Optional[float] = None,
    hierarchy_group_temporal_total_weight: typing.Optional[float] = None,
    window_column: typing.Optional[str] = None,
    window_stride_length: typing.Optional[int] = None,
    window_max_count: typing.Optional[int] = None,
    holiday_regions: typing.Optional[typing.List[str]] = None,
    sync: bool = True,
    create_request_timeout: typing.Optional[float] = None,
    enable_probabilistic_inference: bool = False,
) -> google.cloud.aiplatform.models.Model

Runs the training job and returns a model.

If training on a Vertex AI dataset, you can use one of the following split configurations: Data fraction splits: Any of training_fraction_split, validation_fraction_split and test_fraction_split may optionally be provided, they must sum to up to 1. If the provided ones sum to less than 1, the remainder is assigned to sets as decided by Vertex AI. If none of the fractions are set, by default roughly 80% of data will be used for training, 10% for validation, and 10% for test.

Predefined splits:
Assigns input data to training, validation, and test sets based on the value of a provided key.
If using predefined splits, `predefined_split_column_name` must be provided.
Supported only for tabular Datasets.

Timestamp splits:
Assigns input data to training, validation, and test sets
based on a provided timestamps. The youngest data pieces are
assigned to training set, next to validation set, and the oldest
to the test set.
Supported only for tabular Datasets.

Parameters

Name Description

dataset

datasets.TimeSeriesDataset

Required. The dataset within the same Project from which data will be used to train the Model. The Dataset must use schema compatible with Model being trained, and what is compatible should be described in the used TrainingPipeline's [training_task_definition] [google.cloud.aiplatform.v1beta1.TrainingPipeline.training_task_definition]. For time series Datasets, all their data is exported to training, to pick and choose from.

target_column

str

Required. Name of the column that the Model is to predict values for. This column must be unavailable at forecast.

time_column

str

Required. Name of the column that identifies time order in the time series. This column must be available at forecast.

time_series_identifier_column

str

Required. Name of the column that identifies the time series.

unavailable_at_forecast_columns

List[str]

Required. Column names of columns that are unavailable at forecast. Each column contains information for the given entity (identified by the [time_series_identifier_column]) that is unknown before the forecast (e.g. population of a city in a given year, or weather on a given day).

available_at_forecast_columns

List[str]

Required. Column names of columns that are available at forecast. Each column contains information for the given entity (identified by the [time_series_identifier_column]) that is known at forecast.

forecast_horizon

int

(int): Required. The amount of time into the future for which forecasted values for the target are returned. Expressed in number of units defined by the [data_granularity_unit] and [data_granularity_count] field. Inclusive.

data_granularity_unit

str

Required. The data granularity unit. Accepted values are minute, hour, day, week, month, year.

data_granularity_count

int

Required. The number of data granularity units between data points in the training data. If [data_granularity_unit] is minute, can be 1, 5, 10, 15, or 30. For all other values of [data_granularity_unit], must be 1.

predefined_split_column_name

str

Optional. The key is a name of one of the Dataset's data columns. The value of the key (either the label's value or value in the column) must be one of {TRAIN, VALIDATE, TEST}, and it defines to which set the given piece of data is assigned. If for a piece of data the key is not present or has an invalid value, that piece is ignored by the pipeline. Supported only for tabular and time series Datasets.

timestamp_split_column_name

str

Optional. The key is a name of one of the Dataset's data columns. The value of the key values of the key (the values in the column) must be in RFC 3339 date-time format, where time-offset = "Z" (e.g. 1985-04-12T23:20:50.52Z). If for a piece of data the key is not present or has an invalid value, that piece is ignored by the pipeline. Supported only for tabular and time series Datasets. This parameter must be used with training_fraction_split, validation_fraction_split, and test_fraction_split.

weight_column

str

Optional. Name of the column that should be used as the weight column. Higher values in this column give more importance to the row during Model training. The column must have numeric values between 0 and 10000 inclusively, and 0 value means that the row is ignored. If the weight column field is not set, then all rows are assumed to have equal weight of 1.

time_series_attribute_columns

List[str]

Optional. Column names that should be used as attribute columns. Each column is constant within a time series.

context_window

int

Optional. The amount of time into the past training and prediction data is used for model training and prediction respectively. Expressed in number of units defined by the [data_granularity_unit] and [data_granularity_count] fields. When not provided uses the default value of 0 which means the model sets each series context window to be 0 (also known as "cold start"). Inclusive.

export_evaluated_data_items

bool

Whether to export the test set predictions to a BigQuery table. If False, then the export is not performed.

export_evaluated_data_items_bigquery_destination_uri

string

Optional. URI of desired destination BigQuery table for exported test set predictions. Expected format: bq://<project_id>:<dataset_id>:

If not specified, then results are exported to the following auto-created BigQuery table: <project_id>:export_evaluated_examples_<model_name>_<yyyy_MM_dd'T'HH_mm_ss_SSS'Z'>.evaluated_examples Applies only if [export_evaluated_data_items] is True.

export_evaluated_data_items_override_destination

bool

Whether to override the contents of [export_evaluated_data_items_bigquery_destination_uri], if the table exists, for exported test set predictions. If False, and the table exists, then the training job will fail. Applies only if [export_evaluated_data_items] is True and [export_evaluated_data_items_bigquery_destination_uri] is specified.

quantiles

List[float]

Quantiles to use for the minimize-quantile-loss [AutoMLForecastingTrainingJob.optimization_objective]. This argument is required in this case. Accepts up to 5 quantiles in the form of a double from 0 to 1, exclusive. Each quantile must be unique.

validation_options

str

Validation options for the data validation component. The available options are: "fail-pipeline" - (default), will validate against the validation and fail the pipeline if it fails. "ignore-validation" - ignore the results of the validation and continue the pipeline

budget_milli_node_hours

int

Optional. The train budget of creating this Model, expressed in milli node hours i.e. 1,000 value in this field means 1 node hour. The training cost of the model will not exceed this budget. The final cost will be attempted to be close to the budget, though may end up being (even) noticeably smaller - at the backend's discretion. This especially may happen when further model training ceases to provide any improvements. If the budget is set to a value known to be insufficient to train a Model for the given training set, the training won't be attempted and will error. The minimum value is 1000 and the maximum is 72000.

model_display_name

str

Optional. If the script produces a managed Vertex AI Model. The display name of the Model. The name can be up to 128 characters long and can be consist of any UTF-8 characters. If not provided upon creation, the job's display_name is used.

model_labels

Dict[str, str]

Optional. The labels with user-defined metadata to organize your Models. Label keys and values can be no longer than 64 characters (Unicode codepoints), can only contain lowercase letters, numeric characters, underscores and dashes. International characters are allowed. See https://goo.gl/xmQnxf for more information and examples of labels.

model_id

str

Optional. The ID to use for the Model produced by this job, which will become the final component of the model resource name. This value may be up to 63 characters, and valid characters are [a-z0-9_-]. The first character cannot be a number or hyphen.

parent_model

str

Optional. The resource name or model ID of an existing model. The new model uploaded by this job will be a version of parent_model. Only set this field when training a new version of an existing model.

is_default_version

bool

Optional. When set to True, the newly uploaded model version will automatically have alias "default" included. Subsequent uses of the model produced by this job without a version specified will use this "default" version. When set to False, the "default" alias will not be moved. Actions targeting the model version produced by this job will need to specifically reference this version by ID or alias. New model uploads, i.e. version 1, will always be "default" aliased.

model_version_aliases

Sequence[str]

Optional. User provided version aliases so that the model version uploaded by this job can be referenced via alias instead of auto-generated version ID. A default version alias will be created for the first version of the model. The format is a-z][a-zA-Z0-9-]{0,126}[a-z0-9]

model_version_description

str

Optional. The description of the model version being uploaded by this job.

additional_experiments

List[str]

Optional. Additional experiment flags for the time series forcasting training.

hierarchy_group_columns

List[str]

Optional. A list of time series attribute column names that define the time series hierarchy. Only one level of hierarchy is supported, ex. region for a hierarchy of stores or department for a hierarchy of products. If multiple columns are specified, time series will be grouped by their combined values, ex. (blue, large) for color and size, up to 5 columns are accepted. If no group columns are specified, all time series are considered to be part of the same group.

hierarchy_group_total_weight

float

Optional. The weight of the loss for predictions aggregated over time series in the same hierarchy group.

hierarchy_temporal_total_weight

float

Optional. The weight of the loss for predictions aggregated over the horizon for a single time series.

hierarchy_group_temporal_total_weight

float

Optional. The weight of the loss for predictions aggregated over both the horizon and time series in the same hierarchy group.

window_column

str

Optional. Name of the column that should be used to filter input rows. The column should contain either booleans or string booleans; if the value of the row is True, generate a sliding window from that row.

window_stride_length

int

Optional. Step length used to generate input examples. Every window_stride_length rows will be used to generate a sliding window.

window_max_count

int

Optional. Number of rows that should be used to generate input examples. If the total row count is larger than this number, the input data will be randomly sampled to hit the count.

holiday_regions

List[str]

Optional. The geographical regions to use when creating holiday features. This option is only allowed when data_granularity_unit is day. Acceptable values can come from any of the following levels: Top level: GLOBAL Second level: continental regions NA: North America JAPAC: Japan and Asia Pacific EMEA: Europe, the Middle East and Africa LAC: Latin America and the Caribbean Third level: countries from ISO 3166-1 Country codes.

sync

bool

Whether to execute this method synchronously. If False, this method will be executed in concurrent Future and any downstream object will be immediately returned and synced when the Future has completed.

create_request_timeout

float

Optional. The timeout for the create request in seconds.

enable_probabilistic_inference

bool

If probabilistic inference is enabled, the model will fit a distribution that captures the uncertainty of a prediction. At inference time, the predictive distribution is used to make a point prediction that minimizes the optimization objective. For example, the mean of a predictive distribution is the point prediction that minimizes RMSE loss. If quantiles are specified, then the quantiles of the distribution are also returned. The optimization objective cannot be minimize-quantile-loss.

Exceptions
Type	Description
`RuntimeError`	If Training job has already been run or is waiting to run.

Returns
Type	Description
`model`	The trained Vertex AI Model resource or None if training did not produce a Vertex AI Model.

to_dict

to_dict() -> typing.Dict[str, typing.Any]

Returns the resource proto as a dictionary.

wait

wait()

Helper method that blocks until all futures are complete.

wait_for_resource_creation

wait_for_resource_creation() -> None

Waits until the resource has been created.