ds.monitored_fitdbm: Fits a (multimodal) DBM model
In stefan-m-lenz/dsBoltzmannMachinesClient: Client for dsBoltzmannMachines

The procedure for DBM fitting consists of two parts: First a stack of RBMs is pretrained in a greedy layerwise manner (see ds.monitored_stackrbms). Then the weights of all layers are jointly trained using the general Boltzmann machine learning procedure. During pre-training and fine-tuning, monitoring data is collected by default. The monitoring data is returned to the user. The trained model is stored on the server side (see parameter newobj).

ds.monitored_fitdbm(
  datasources,
  newobj = "dbm",
  data = "D",
  monitoring = "logproblowerbound",
  monitoringdata = data,
  monitoringpretraining = "reconstructionerror",
  monitoringdatapretraining = monitoringdata,
  nhiddens = NULL,
  epochs = NULL,
  nparticles = NULL,
  learningrate = NULL,
  learningrates = NULL,
  learningratepretraining = NULL,
  epochspretraining = NULL,
  batchsizepretraining = NULL,
  pretraining = NULL
)

`datasources`	A list of Opal object(s) as a handle to the server-side session
`newobj`	The name of the variable in which the trained DBM will be stored. Defaults to `"dbm"`
`data`	The name of the variable that holds the data on the server-side. Defaults to `"D"`.
`monitoring`	Name(s) for the monitoring options used for monitoring the fine-tuning. Possible options: `"logproblowerbound"`: Variational lower bound of log probability (Default) `"exactloglikelihood"`: Exact calculation of log-likelihood. This is only feasible for very small models. `NULL`: No monitoring
`monitoringdata`	A vector of names for server-side data sets that are to be used for monitoring
`monitoringpretraining`	Name for monitoring options used for monitoring the pre-training. The options are the same as for training an RBM (see `ds.monitored_fitrbm`). By default, the reconstruction error is monitored.
`monitoringdatapretraining`	A vector of names for data sets that are to be used for monitoring the pretraining. By default, this is the same as the `monitoringdata`.
`nhiddens`	A vector that defines the number of nodes in the hidden layers of the DBM. The default value specifies two hidden layers with the same size as the visible layer.
`epochs`	Number of training epochs for fine-tuning, defaults to 10
`nparticles`	Number of particles used for sampling during fine-tuning of the DBM, defaults to 100
`learningrate`	Learning rate for joint training of layers (= fine-tuning) using the learning algorithm for a general Boltzmann machine with mean-field approximation. The learning rate for fine tuning is by default decaying with the number of epochs, starting with the given value for the `learningrate`. By default, the learning rate decreases with the factor 11 / (10 + epoch).
`learningrates`	A vector of learning rates for each epoch of fine-tuning
`learningratepretraining`	Learning rate for pretraining, defaults to `learningrate`
`epochspretraining`	Number of training epochs for pretraining, defaults to `epochs`
`batchsizepretraining`	Batchsize for pretraining, defaults to 1
`pretraining`	The arguments for layerwise pretraining can be specified for each layer individually. This is done via a vector of names for objects that have previously been defined by `ds.bm.defineLayer` or `ds.bm.definePartitionedLayer`. (For a detailed description of the possible parameters, see the help there). If the number of training epochs and the learning rate are not specified explicitly for a layer, the values of `epochspretraining`, `learningratepretraining` and `batchsizepretraining` are used.

If the option dsBoltzmannMachines.shareModels is set to TRUE by an administrator at the server side, the models themselves are returned in addition.

stefan-m-lenz/dsBoltzmannMachinesClient documentation built on May 2, 2021, 10:46 a.m.