Skip to main content

Table 1 Most important hyperparameters for fitting a DBM. These parameters can be specified in the function “ds.monitored_fitdbm” (see Table 2). The parameters for pre-training can also be controlled individually for each layer (i.e. for each RBM in the stack) via the function “ds.bm.defineLayer”. Together with the function “ds.bm.definePartitionedLayer”, this allows to also create models with partitioned architectures

From: Deep generative models in DataSHIELD

Hyperparameter name Meaning of hyperparameter
learningrate Learning rate for stochastic gradient descent optimization
learningratepretraining Learning rate for pre-training, may be specified separately
epochs Number of training epochs
epochspretraining Number of epochs for pre-training, may be specified separately
nhiddens Number of hidden nodes specified as a vector of numbers, containing one number for each hidden layer
batchsizepretraining Batch size used in pre-training