Struct LdaParams 

pub struct LdaParams<F: BixverseFloat> {
    pub alpha: F,
    pub alpha_by_topic: bool,
    pub eta: F,
    pub eta_by_topic: bool,
    pub max_iter: usize,
    pub tol: F,
    pub inner_max_iter: usize,
    pub inner_tol: F,
    pub check_every: usize,
    pub learning: LdaLearning,
    pub seed: u64,
}

Options for the variational Bayes LDA solver.

Defaults follow pycisTopic so the knobs mean the same thing on both sides.

Fields

alpha: F

Dirichlet prior on the document-topic distributions.

alpha_by_topic: bool

Whether alpha is divided by the topic count. true gives the Griffiths and Steyvers 50 / k heuristic that cisTopic defaults to.

eta: F

Dirichlet prior on the topic-term distributions.

eta_by_topic: bool

Whether eta is divided by the topic count.

max_iter: usize

Maximum outer iterations (batch), ignored by the online variant which counts epochs instead.

tol: F

Relative change in the bound below which the solver stops.

inner_max_iter: usize

Maximum fixed-point iterations of the per-document E-step.

inner_tol: F

Relative L1 change in gamma below which the per-document E-step stops. Relative rather than absolute, see [vb::e_step_document].

check_every: usize

Cadence, in iterations, for evaluating the bound and testing tol.

learning: LdaLearning

How the corpus is swept, see LdaLearning.

seed: u64

Seed for the lambda and gamma initialisation.

Implementations

impl<F: BixverseFloat> LdaParams<F>

pub fn new( alpha: F, alpha_by_topic: bool, eta: F, eta_by_topic: bool, max_iter: usize, tol: F, inner_max_iter: usize, inner_tol: F, check_every: usize, learning: LdaLearning, seed: u64, ) -> Self

Generate a new instance of LdaParams

Params

• alpha - Document-topic Dirichlet prior.
• alpha_by_topic - Divide alpha by the topic count.
• eta - Topic-term Dirichlet prior.
• eta_by_topic - Divide eta by the topic count.
• max_iter - Maximum outer iterations.
• tol - Relative bound change for convergence.
• inner_max_iter - Maximum per-document E-step iterations.
• inner_tol - Relative L1 gamma change for E-step convergence.
• check_every - Iterations between bound evaluations.
• learning - Batch or online, see LdaLearning.
• seed - Seed for the initialisation.

Returns

The initialised LdaParams.

pub fn resolve_priors(&self, k: usize) -> Result<(F, F), MethodsErrors>

Resolve the effective Dirichlet priors for a given topic count.

Params

• k - Number of topics.

Returns

The (alpha, eta) actually handed to the solver, or an error if either is not strictly positive.

Trait Implementations

impl<F: Clone + BixverseFloat> Clone for LdaParams<F>

fn clone(&self) -> Self

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl<F: Copy + BixverseFloat> Copy for LdaParams<F>

impl<F: Debug + BixverseFloat> Debug for LdaParams<F>

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl<F: BixverseFloat> Default for LdaParams<F>

fn default() -> Self

Returns the “default value” for a type.