tfr.keras.pipeline.PipelineHparams

Hyperparameters used in ModelFitPipeline.

tfr.keras.pipeline.PipelineHparams(
    model_dir: str,
    num_epochs: int,
    steps_per_epoch: int,
    validation_steps: int,
    learning_rate: float,
    loss: Union[str, Dict[str, str]],
    loss_reduction: str = tf.losses.Reduction.AUTO,
    optimizer: str = 'adam',
    loss_weights: Optional[Union[float, Dict[str, float]]] = None,
    steps_per_execution: int = 10,
    automatic_reduce_lr: bool = False,
    early_stopping_patience: int = 0,
    early_stopping_min_delta: float = 0.0,
    use_weighted_metrics: bool = False,
    export_best_model: bool = False,
    best_exporter_metric_higher_better: bool = False,
    best_exporter_metric: str = 'loss',
    strategy: Optional[str] = None,
    cluster_resolver: Optional[tf.distribute.cluster_resolver.ClusterResolver] = None,
    variable_partitioner: Optional[tf.distribute.experimental.partitioners.Partitioner] = None,
    tpu: Optional[str] = ''
)

Hyperparameters to be specified for ranking pipeline.

model_dir A path to output the model and training data.
num_epochs An integer to specify the number of epochs of training.
steps_per_epoch An integer to specify the number of steps per epoch. When it is None, going over the training data once is counted as an epoch.
validation_steps An integer to specify the number of validation steps in each epoch. Note that a mini-batch of data will be evaluated in each step and this is the number of steps taken for validation in each epoch.
learning_rate A float to indicate the learning rate of the optimizer.
loss A string or a map to strings that indicate the loss to be used. When loss is a string, all outputs and labels will be trained with the same loss. When loss is a map, outputs and labels will be trained with losses implied by the corresponding keys.
loss_reduction An option in tf.keras.losses.Reduction to specify the reduction method.
optimizer An option in tf.keras.optimizers identifiers to specify the optimizer to be used.
loss_weights None or a float or a map to floats that indicate the relative weights for each loss. When not specified, all losses are applied with the same weight 1.
steps_per_execution An integer to specify the number of steps executed in each operation. Tuning this to optimize the training performance in distributed training.
automatic_reduce_lr A boolean to indicate whether to use ReduceLROnPlateau callback.
early_stopping_patience Number of epochs with no improvement after which training will be stopped.
early_stopping_min_delta Minimum change in the monitored quantity to qualify as an improvement, i.e. an absolute change of less than early_stopping_min_delta, will count as no improvement.
use_weighted_metrics A boolean to indicate whether to use weighted metrics.
export_best_model A boolean to indicate whether to export the best model evaluated by the best_exporter_metric on the validation data.
best_exporter_metric_higher_better A boolean to indicate whether the best_exporter_metric is the higher the better.
best_exporter_metric A string to specify the metric used to monitor the training and to export the best model. Default to the 'loss'.
strategy An option of strategies supported in strategy_utils. Choose from ["MirroredStrategy", "MultiWorkerMirroredStrategy", "ParameterServerStrategy", "TPUStrategy"].
cluster_resolver A cluster_resolver to build strategy.
variable_partitioner Variable partitioner to be used in ParameterServerStrategy.
tpu TPU address for TPUStrategy. Not used for other strategy.

Methods

__eq__

__eq__(
    other
)

automatic_reduce_lr False
best_exporter_metric 'loss'
best_exporter_metric_higher_better False
cluster_resolver None
early_stopping_min_delta 0.0
early_stopping_patience 0
export_best_model False
loss_reduction 'auto'
loss_weights None
optimizer 'adam'
steps_per_execution 10
strategy None
tpu ''
use_weighted_metrics False
variable_partitioner None