Models and scores

Availiable models:

• BaseModel — Parent class for model creation
• TopicModel — a wrapper class for bigartm topic model
• DummyTopicModel — a fake model that contains training information but not actual artm model. Needed to save memory space during the training.

Availiable scores:

• BaseScore — a parent class for all the Strategies
• ExampleScore — Example of minimal working example of custom score
• IntratextCoherenceScore — score that calculates coherence as a measure of interpretability of the model using raw documents from dataset. Calculation-heavy score. Recommended to be used after model training
• BleiLaffertyScore — An experimental light-weight score to estimate interpretability of the topics
• SemanticRadiusScore — An experimental score reflecting whether collection is adequately described by topics. Lower if better. Calculation-heavy score.

Internal model structure

main model attributes:

• model_id — a model string id, unique for its Experiment.

• scores — dict of lists, each list corresponds to the score value or list of values at certain training stage.

• custom_scores — variable providing custom scores for the model

• custom_regularizers — variable providing custom regularizers for the model. An example is provided in topic_prior_regularizer.py.

main model methods:

• _fit — function performing model training. Takes the dataset and number of iterations. Optionally, you can pass custom_regularizers here, if you wish to apply them to a single iteration.

  Important Notice! We assume that the model training happens through Cube interface and this method, while important should never be used by users if they are hope to have their actions logged

• get_phi — function that returns p(token|topic/cluster) probability distributions that returns pandas.DataFrame with tokens as index and topics/clusters as columns

  Important Notice! Strictly speaking the function returns degree to which token belongs to the topic/cluster and shouldn’t be a probability distribution. But scince its main use-case intended for topic models some of the functions using this method might work incorrectly in non-distribution case

• get_theta — function that returns p(topic/cluster|document) probability distributions that returns pandas.DataFrame with topics/clusters as index and document ids as columns.

  Important Notice! Strictly speaking the function returns degree to which document belongs to the topic/cluster and shouldn’t be a probability distribution. But scince its main use-case intended for topic models some of the functions using this method might work incorrectly in non-distribution case

• save — saves model to the path directory.

• load — loads model from the path directory

• clone — creates copy of a model.

• get_jsonable_from_parameters — turns model parameters to jsonable format for logging purposes

What do you need to create your own model?

Following this steps you should be able to code a model integrated with the library methods:

1. New model class is inherrited from BaseModel

2. A child class should contain methods __init__, _fit, get_phi, get_theta, save, load, clone, get_jsonable_from_parameters.

from .base_model import BaseModel
from .topic_model import TopicModel
from .dummy_topic_model import DummyTopicModel

from .base_score import BaseScore
from .example_score import ScoreExample
from .intratext_coherence_score import IntratextCoherenceScore

SUPPORTED_MODEL_CLASSES = (
    TopicModel,
)