train_rfn: Trains a Rectified Factor Network (RFN)
In bioinf-jku/librfn-R: Rectified Factor Networks

Description Usage Arguments Value

Trains an RFN as described by Clevert et al., 2014

train_rfn(X, n_hidden, n_iter, etaW, etaP, minP, batch_size = -1,
  dropout_rate = 0, input_noise_rate = 0, l2_weightdecay = 0,
  l1_weightdecay = 0, h_threshold = 0, momentum = 0,
  noise_type = "saltpepper", activation = "relu", apply_scaling = 1,
  apply_newton_update = 1, seed = -1, use_gpu = 0, gpu_id = -1,
  verbose = T)

`X`	The data matrix
`n_hidden`	Number of latent variables to estimate
`n_iter`	Number of iterations to run the algorithm
`etaW`	Learning rate of the W parameter
`etaP`	Learning rate of the Psi parameter (It's probably save to set this to the same value as etaW)
`minP`	Minimal value for Psi. Should be in 1e-5 - 1e-1
`batch_size`	If > 2, this will activate mini-batch learning instead of full batch learning
`dropout_rate`	Dropout rate for the latent variables
`input_noise_rate`	Noise/dropout rate for input variables
`l2_weightdecay`	L2 penalty for weight decay
`l1_weightdecay`	L1 penalty for weight decay
`h_threshold`	Threshhold for rectifying/leaky activations
`momentum`	Momentum term for learning
`noise_type`	Type of input noise. One of "dropout", "saltpepper", "gaussian"
`activation`	Activation function for hidden/latent variables. One of "linear", "relu", "leaky", "sigmoid", "tanh"
`apply_scaling`	Scale the data
`apply_newton_update`	Whether to use a Newton update (default) or a gradient descent step
`seed`	Seed for the random number generator
`use_gpu`	Use the gpu (default cpu). Works only for sparse input.
`gpu_id`	If use_gpu is true, use gpu with this id (default -1 selects one available)
`verbose`	True to print messages during the run. Default is True.

Returns a list of matrices W, P, H, Wout, whereas W %*% H is the noise-free reconstruction of the data X and diag(P) is the covariance matrix of the additive zero-mean noise. Wout is the matrix that maps input vectors to their latent representation, usually by pmax(t(Wout) %*% X, 0).