Skip to contents

Wrapper function for the LDA parameters

params_lda.Rd

Solver options for the variational Bayes latent Dirichlet allocation, see run_lda().

Usage

params_lda(
  alpha = 50,
  alpha_by_topic = TRUE,
  eta = 0.1,
  eta_by_topic = FALSE,
  max_iter = 150L,
  tol = 0.001,
  inner_max_iter = 100L,
  inner_tol = 0.001,
  check_every = 10L,
  learning = c("batch", "online"),
  batch_size = 1024L,
  n_epochs = 10L
)

Arguments

alpha

Numeric. Dirichlet prior on the document-topic distributions. Defaults to 50.0.

alpha_by_topic

Boolean. Shall alpha be divided by the topic count. Defaults to TRUE.

eta

Numeric. Dirichlet prior on the topic-term distributions. Defaults to 0.1.

eta_by_topic

Boolean. Shall eta be divided by the topic count. Defaults to FALSE.

max_iter

Integer. Maximum outer iterations. Ignored by the online variant, which counts epochs instead. Defaults to 150L.

tol

Numeric. Relative change in the bound below which the solver stops. Defaults to 0.001.

inner_max_iter

Integer. Maximum fixed-point iterations of the per-document E-step. Defaults to 100L.

inner_tol

Numeric. Relative L1 change in the variational parameters below which the per-document E-step stops. Defaults to 0.001.

check_every

Integer. Iterations between bound evaluations. Defaults to 10L.

learning

String. Batch or online variational inference. One of c("batch", "online"). Defaults to "batch".

batch_size

Integer. Documents per mini-batch. Online only. Defaults to 1024L.

n_epochs

Integer. Passes over the corpus. Online only. Defaults to 10L.

Value

A named list with the following elements:

  • alpha - Numeric. Dirichlet prior on the document-topic distributions. Defaults to 50.0.

  • alpha_by_topic - Boolean. Shall alpha be divided by the topic count. Defaults to TRUE.

  • eta - Numeric. Dirichlet prior on the topic-term distributions. Defaults to 0.1.

  • eta_by_topic - Boolean. Shall eta be divided by the topic count. Defaults to FALSE.

  • max_iter - Integer. Maximum outer iterations. Ignored by the online variant, which counts epochs instead. Defaults to 150L.

  • tol - Numeric. Relative change in the bound below which the solver stops. Defaults to 0.001.

  • inner_max_iter - Integer. Maximum fixed-point iterations of the per-document E-step. Defaults to 100L.

  • inner_tol - Numeric. Relative L1 change in the variational parameters below which the per-document E-step stops. Defaults to 0.001.

  • check_every - Integer. Iterations between bound evaluations. Defaults to 10L.

  • learning - String. Batch or online variational inference. One of c("batch", "online"). Defaults to "batch".

  • batch_size - Integer. Documents per mini-batch. Online only. Defaults to 1024L.

  • n_epochs - Integer. Passes over the corpus. Online only. Defaults to 10L.

Details

The defaults follow pycisTopic, so the knobs mean the same thing on both sides. alpha_by_topic = TRUE turns alpha into the Griffiths and Steyvers 50 / k heuristic that cisTopic defaults to; set it to FALSE if you want alpha taken literally.

learning = "batch" sweeps every document once per iteration and is monotone in the bound. "online" takes decaying steps from shuffled mini-batches, which reaches a usable fit in far fewer passes on a large corpus at the cost of that guarantee. batch_size and n_epochs are only read by the online variant.

References

Hoffman, Blei and Bach, NIPS, 2010; Bravo Gonzalez-Blas, et al., Nat Methods, 2019