kumoai.trainer.OptimizationPlan#
- class kumoai.trainer.OptimizationPlan[source]#
Configuration parameters that define how columns are encoded in the training and batch prediction pipelines. Please see the Kumo documentation for more information.
- Variables:
max_epochs – (
int
) The maximum number of epochs to train a model for (default:run_mode
-dependent). Supported Task Types: Allmin_steps_per_epoch – (
int
) The minimum number of steps to be included in an epoch; one step corresponds to one forward pass of a mini-batch (default:30
). Supported Task Types: Allmax_steps_per_epoch – (
int
) The maximum number of steps to be included in an epoch; one step corresponds to one forward pass of a mini-batch (default:run_mode
-dependent). Supported Task Types: Allmax_val_steps – (
int
) The maximum number of steps to be included in a validation pass; one step corresponds to one forward pass of a mini-batch (default:run_mode
-dependent). Supported Task Types: Allmax_test_steps – (
int
) The maximum number of steps to be included in a test pass; one step corresponds to one forward pass of a mini-batch (default:run_mode
-dependent). Supported Task Types: Allloss – (
list[str]
) The loss type to use in the model optimizer (default:task_type
-dependent). Supported Task Types: Allbase_lr – (
list[float]
) The base learning rate (pre-decay) to be used in the model optimizer. (default:[1e-4, 5e-4, 1e-3, 5e-3, 1e-2, 5e-2]
). Supported Task Types: Allweight_decay – (
list[float]
) A list of potential weight decay options in the model optimizer. (default:[0.0, 5e-8, 5e-7, 5e-6]
). Supported Task Types: Allbatch_size – (
list[int]
) The number of examples to be included in one mini-batch. (default:[512, 1024]
). Supported Task Types: Allearly_stopping – (
list[EarlyStoppingConfig]
) A list of potential early stopping strategiesEarlyStoppingConfig
for model optimization (default:[{min_delta=0.0, patience=3}]
). Supported Task Types: Alllr_scheduler – (
list[LRSchedulerConfig]
) A list of potential learning rate schedulersLRSchedulerConfig
for model optimization (default:[ {name="cosine_with_warmup_restarts", interval="step"}, {name="constant_with_warmup", interval="step"}, {name="linear_with_warmup", interval="step"}, {name="csoine_with_warmup", interval="step"}]
). Supported Task Types: Allmajority_sampling_ratio – (
list[float | None]
) A ratio to specify how examples are sampled from the majority class (default:[None]
). Supported Task Types: Binary Classificationweight_mode – (
list[WeightMode | None]
) Defines how to use the weight column in the training table (if present) in the model training process. If [None] weight column will not be used for training even if present. It could still be used for metrics computation.target_fraction – (
list[float]
) Specifies the fraction of ground-truth links to be used for supervision during training. By default, all ground-truth links are included (1.0
). Setting a lower value limits the number of links used for supervision, which can help prevent all existing message-passing edges from being supervised. Regardless of the value specified, at least one edge per entity is always retained for supervision to ensure adequate coverage. Supported Task Types: Static Link Predictionoverlap_ratio – (
list[float]
) Denotes the overlap ratio between message passing edges and remaining supervision edges according totarget_fraction
. By default, all remaining ground-truth links will also be used during message passing (1.0
). A lower value prevents that message passing edges and supervision edges will be shared during training. Supported Task Types: Static Link Prediction