Skip to content

Latest commit

 

History

History
89 lines (63 loc) · 3.06 KB

sampler.md

File metadata and controls

89 lines (63 loc) · 3.06 KB
Raw
layout title
default
High-speed Sampler

The DimmWitted high-speed sampler

This document briefly presents DimmWitted, a high-speed Gibbs sampler for DeepDive.

In deepdive.conf, you can change the sampler executable as follows:

deepdive {
  sampler.sampler_cmd: "util/sampler-dw-mac gibbs"
}

Use sampler-dw-mac or sampler-dw-linux depending on which type of system your are on.

Since version 0.03, DeepDive automatically chooses the correct executable based on the system environment, so we recommend to omit the sampler_cmd directive.

Sampler arguments

The sampler executable can be invoked independently of DeepDive. The following arguments to the sampler executable can be used:

-q, --quiet
    Quiet output

-c <int>,  --n_datacopy <int> (Linux only)
    Number of data copies. Each NUMA node has a copy of factor graph. This
    argument specifies number of NUMA nodes to use. Default is using all
    NUMA nodes.

-w <weightsFile> | --weights <weightsFile>
    Weights file (required)
    It is a binary format file output by DeepDive.

-v <variablesFile> | --variables <variablesFile>
    Variables file (required)
    It is a binary format file output by DeepDive.

-f <factorsFile> | --factors <factorsFile>
    Factors file (required)
    It is a binary format file output by DeepDive.

-m <metaFile> | --fg_meta <metaFile>
    Factor graph meta data file file (required)
    It is a text file containing factor graph meta information
    as well as paths to weight/variable/factor/edge files.

-o <outputFile> | --outputFile <outputFile>
    Output file path (required)

-i <numSamplesInference> | --n_inference_epoch <numSamplesInference>
    Number of iterations (epochs) during inference (required)

-l <learningNumIterations> | --n_learning_epoch <learningNumIterations>
    Number of iterations (epochs) during weight learning (required)

-s <learningNumSamples> | --n_samples_per_learning_epoch <learningNumSamples>
    Number of samples per iteration during weight learning (required)

-a <learningRate> | --alpha <learningRate> | --stepsize <learningRate>
    The learning rate for gradient descent (default: 0.1)

-d <diminishRate> | --diminish <diminishRate>
    The diminish rate for learning (default: 0.95).
    Learning rate will shrink by this parameter after each iteration.

-b <regularizationParameter> | --reg_param <regularizationParameter>
    The l2 regularization parameter for learning (default: 0.01).

--sample_evidence
    Output probablities for evidence variables. Default is off, i.e., output
    only contains probabilities for non-evidence variables.

--learn_non_evidence
    Sample non-evidence variables during learning. Default if off. This option
    should be turned on if there exists a factor connecting evidence and non-evidence
    variables.

You can see a detailed list by running deepdive env sampler-dw gibbs --help.