ML2Pvae (version 1.0.0.1)

build_vae_independent: Build a VAE that fits to a standard N(0,I) latent distribution with independent latent traits

Description

Build a VAE that fits to a standard N(0,I) latent distribution with independent latent traits

Usage

build_vae_independent(
  num_items,
  num_skills,
  Q_matrix,
  model_type = 2,
  enc_hid_arch = c(ceiling((num_items + num_skills)/2)),
  hid_enc_activations = rep("sigmoid", length(enc_hid_arch)),
  output_activation = "sigmoid",
  kl_weight = 1,
  learning_rate = 0.001
)

Arguments

num_items

an integer giving the number of items on the assessment; also the number of nodes in the input/output layers of the VAE

num_skills

an integer giving the number of skills being evaluated; also the dimensionality of the distribution learned by the VAE

Q_matrix

a binary, num_skills by num_items matrix relating the assessment items with skills

model_type

either 1 or 2, specifying a 1 parameter (1PL) or 2 parameter (2PL) model; if 1PL, then all decoder weights are fixed to be equal to one

enc_hid_arch

a vector detailing the size of hidden layers in the encoder; the number of hidden layers is determined by the length of this vector

hid_enc_activations

a vector specifying the activation function in each hidden layer in the encoder; must be the same length as enc_hid_arch

output_activation

a string specifying the activation function in the output of the decoder; the ML2P model always uses 'sigmoid'

kl_weight

an optional weight for the KL divergence term in the loss function

learning_rate

an optional parameter for the adam optimizer

Value

returns three keras models: the encoder, decoder, and vae.

Examples

# NOT RUN {
Q <- matrix(c(1,0,1,1,0,1,1,0), nrow = 2, ncol = 4)
models <- build_vae_independent(4, 2, Q,
          enc_hid_arch = c(6, 3), hid_enc_activation = c('sigmoid', 'relu'),
          output_activation = 'tanh', kl_weight = 0.1)
models <- build_vae_independent(4, 2, Q)
vae <- models[[3]]
# }