SageMaker smart sifting Python SDK reference - Amazon SageMaker AI
SageMaker smart sifting configuration modulesSageMaker smart sifting data batch transform modulesSageMaker smart sifting loss implementation moduleSageMaker smart sifting data loader wrapper module

SageMaker smart sifting Python SDK reference

This page provides a reference of Python modules you need for applying SageMaker smart sifting to your training script.

SageMaker smart sifting configuration modules

class smart_sifting.sift_config.sift_configs.RelativeProbabilisticSiftConfig()

The SageMaker smart sifting configuration class.

Parameters

  • beta_value (float) – A beta (constant) value. It is used to calculate the probability of selecting a sample for training based on the percentile of the loss in the loss values history. Lowering the beta value results in a lower percentage of data sifted, and raising it results in a higher percentage of data sifted. There’s no minimum or maximum value for the beta value, other than it must be a positive value. The following reference table gives information for sifting rates with respect to beta_value.

    beta_value Proportion of data kept (%) Proportion of data sifted out (%)
    0.1 90.91 9.01
    0.25 80 20
    0.5 66.67 33.33
    1 50 50
    2 33.33 66.67
    3 25 75
    10 9.09 90.92
    100 0.99 99.01

  • loss_history_length (int) – The number of previous training losses to store for the relative threshold loss based sampling.

  • loss_based_sift_config (dict or a LossConfig object) – Specify a LossConfig object that returns the SageMaker smart sifting Loss interface configuration.

class smart_sifting.sift_config.sift_configs.LossConfig()

The configuration class for the loss_based_sift_config parameter of the RelativeProbabilisticSiftConfig class.

Parameters

  • sift_config (dict or a SiftingBaseConfig object) – Specify a SiftingBaseConfig object that returns a sifting base configuration dictionary.

class smart_sifting.sift_config.sift_configs.SiftingBaseConfig()

The configuration class for the sift_config parameter of LossConfig.

Parameters

  • sift_delay (int) – The number of training steps to wait for before starting sifting. We recommend that you start sifting after all the layers in the model have enough view of the training data. The default value is 1000.

  • repeat_delay_per_epoch (bool) – Specify whether to delay sifting every epoch. The default value is False.

SageMaker smart sifting data batch transform modules

class smart_sifting.data_model.data_model_interface.SiftingBatchTransform

A SageMaker smart sifting Python module for defining how to perform batch transform. Using this, you can set up a batch transform class that converts the data format of your training data to SiftingBatch format. SageMaker smart sifting can sift and accumulate data in this format into a sifted batch.

class smart_sifting.data_model.data_model_interface.SiftingBatch

An interface to define a batch data type that can be sifted and accumulated.

class smart_sifting.data_model.list_batch.ListBatch

A module for keeping track of a list batch for sifting.

class smart_sifting.data_model.tensor_batch.TensorBatch

A module for keeping track of a tensor batch for sifting.

SageMaker smart sifting loss implementation module

class smart_sifting.loss.abstract_sift_loss_module.Loss

A wrapper module for registering the SageMaker smart sifting interface to the loss function of a PyTorch-based model.

SageMaker smart sifting data loader wrapper module

class smart_sifting.dataloader.sift_dataloader.SiftingDataloader

A wrapper module for registering the SageMaker smart sifting interface to the data loader of a PyTorch-based model.

The Main Sifting Dataloader iterator sifts out training samples from a dataloader based on a sift configuration.

Parameters

  • sift_config (dict or a RelativeProbabilisticSiftConfig object) – A RelativeProbabilisticSiftConfig object.

  • orig_dataloader (a PyTorch DataLoader object) – Specify the PyTorch Dataloader object to be wrapped.

  • batch_transforms (a SiftingBatchTransform object) – (Optional) If your data format is not supported by the SageMaker smart sifting library’s default transform, you must create a batch transform class using the SiftingBatchTransform module. This parameter is used to pass the batch transform class. This class is used for SiftingDataloader to convert the data into a format that the SageMaker smart sifting algorithm can accept.

  • model (a PyTorch model object) – The original PyTorch model

  • loss_impl (a sifting loss function of smart_sifting.loss.abstract_sift_loss_module.Loss) – A sifting loss function that is configured with the Loss module and wraps the PyTorch loss function.

  • log_batch_data (bool) – Specify whether to log batch data. If set to True, SageMaker smart sifting logs the details of the batches that are kept or sifted. We recommend that you turn it on only for a pilot training job. When logging is on, the samples are loaded to GPU and transferred to CPU, which introduces overhead. The default value is False.