TABULARGAN Procedure

TRAIN Statement

TRAIN <options>;

The TRAIN statement specifies the training parameters to use for training the tabular GAN model.

Table 3 summarizes the options in this statement.

Table 3: TRAIN Statement Options

Option	Description
CHECKPOINTFREQ=	Specifies the frequency at which to save a checkpoint during training
CHECKPOINTIN=	Specifies the checkpoint table to be loaded to restore your model
CHECKPOINTOUT=	Specifies the table in which to save the current model as a checkpoint
EMBEDDINGDIM=	Specifies the dimensions of the input noise to the generator
MINIBATCHSIZE=	Specifies the number of observations in one minibatch
PACKSIZE=	Specifies the number of samples to group together in applying the discriminator
PRINTFREQ=	Specifies the frequency at which to print losses to the log
REGR1WEIGHT=	Specifies the weight to use in regularizing the discriminator
USEORIGLEVELFREQ	Uses the log-frequency of categorical levels in the conditional sampling

You can specify the following options:

CHECKPOINTFREQ=number

specifies the frequency at which to save a checkpoint during training.

By default, CHECKPOINTFREQ=5000.

CHECKPOINTIN=libref.data-table

specifies the checkpoint table to be loaded to restore your model. If you omit this parameter, the model is trained from scratch. libref.data-table is a two-level name, where libref refers to the library, and data-table specifies the name of the output data table. For more information about this two-level name, see the DATA= option and the section Using SAS Viya Workbench.

CHECKPOINTOUT=libref.data-table

specifies the table in which to save the current model as a checkpoint. libref.data-table is a two-level name, where libref refers to the library, and data-table specifies the name of the output data table. For more information about this two-level name, see the DATA= option and the section Using SAS Viya Workbench.

EMBEDDINGDIM=number

specifies the dimensions of the input noise to the generator.

By default, EMBEDDINGDIM=128.

MINIBATCHSIZE=number

specifies the number of observations in one minibatch. The number should be smaller than the total number of observations in the input data set.

By default, MINIBATCHSIZE=500.

PACKSIZE=number

specifies the number of samples to group together in applying the discriminator. The number must be a factor of the value that you specify in the MINIBATCHSIZE= option.

By default, PACKSIZE=10.

PRINTFREQ=number

specifies the frequency at which to print losses to the log and the iteration history table.

By default, PRINTFREQ=1.

REGR1WEIGHT=number

specifies the weight to use in regularizing the discriminator.

By default, REGR1WEIGHT=10.

USEORIGLEVELFREQ

uses the log-frequency of categorical levels in the conditional sampling.

Last updated: September 23, 2026