models.callbacks – Callbacks for track and viz LDA train process

Callbacks can be used to observe the training process.

Since training in huge corpora can be time consuming, we want to offer the users some insight into the process, in real time. In this way, convergence issues or other potential problems can be identified early in the process, saving precious time and resources.

The metrics exposed through this module can be used to construct Callbacks, which will be called at specific points in the training process, such as “epoch starts” or “epoch finished”. These metrics can be used to assess mod’s convergence or correctness, for example to save the model, visualize intermediate results, or anything else.

Usage examples

To implement a Callback, inherit from this base class and override one or more of its methods.

Create a callback to save the training model after each epoch

>>> from gensim.test.utils import get_tmpfile
>>> from gensim.models.callbacks import CallbackAny2Vec
>>>
>>>
>>> class EpochSaver(CallbackAny2Vec):
...     '''Callback to save model after each epoch.'''
...
...     def __init__(self, path_prefix):
...         self.path_prefix = path_prefix
...         self.epoch = 0
...
...     def on_epoch_end(self, model):
...         output_path = get_tmpfile('{}_epoch{}.model'.format(self.path_prefix, self.epoch))
...         model.save(output_path)
...         self.epoch += 1
...

Create a callback to print progress information to the console:

>>> class EpochLogger(CallbackAny2Vec):
...     '''Callback to log information about training'''
...
...     def __init__(self):
...         self.epoch = 0
...
...     def on_epoch_begin(self, model):
...         print("Epoch #{} start".format(self.epoch))
...
...     def on_epoch_end(self, model):
...         print("Epoch #{} end".format(self.epoch))
...         self.epoch += 1
...
>>>
>>> epoch_logger = EpochLogger()
>>> w2v_model = Word2Vec(common_texts, iter=5, size=10, min_count=0, seed=42, callbacks=[epoch_logger])
Epoch #0 start
Epoch #0 end
Epoch #1 start
Epoch #1 end
Epoch #2 start
Epoch #2 end
Epoch #3 start
Epoch #3 end
Epoch #4 start
Epoch #4 end

Create and bind a callback to a topic model. This callback will log the perplexity metric in real time:

>>> from gensim.models.callbacks import PerplexityMetric
>>> from gensim.models.ldamodel import LdaModel
>>> from gensim.test.utils import common_corpus, common_dictionary
>>>
>>> # Log the perplexity score at the end of each epoch.
>>> perplexity_logger = PerplexityMetric(corpus=common_corpus, logger='shell')
>>> lda = LdaModel(common_corpus, id2word=common_dictionary, num_topics=5, callbacks=[perplexity_logger])
class gensim.models.callbacks.Callback(metrics)

Bases: object

A class representing routines called reactively at specific phases during trained.

These can be used to log or visualize the training progress using any of the metric scores developed before. The values are stored at the end of each training epoch. The following metric scores are currently available:

Parameters

metrics (list of Metric) – The list of metrics to be reported by the callback.

on_epoch_end(epoch, topics=None)

Report the current epoch’s metric value.

Called at the end of each training iteration.

Parameters
  • epoch (int) – The epoch that just ended.

  • topics (list of list of str, optional) – List of tokenized topics. This is required for the coherence metric.

Returns

Mapping from metric names to their values. The type of each value depends on the metric type, for example DiffMetric computes a matrix while ConvergenceMetric computes a float.

Return type

dict of (str, object)

set_model(model)

Save the model instance and initialize any required variables which would be updated throughout training.

Parameters

model (BaseTopicModel) – The model for which the training will be reported (logged or visualized) by the callback.

class gensim.models.callbacks.CallbackAny2Vec

Bases: object

Base class to build callbacks for Word2Vec & subclasses.

Callbacks are used to apply custom functions over the model at specific points during training (epoch start, batch end etc.). This is a base class and its purpose is to be inherited by custom Callbacks that implement one or more of its methods (depending on the point during training where they want some action to be taken).

See examples at the module level docstring for how to define your own callbacks by inheriting from this class.

on_batch_begin(model)

Method called at the start of each batch.

Parameters

model (Word2Vec or subclass) – Current model.

on_batch_end(model)

Method called at the end of each batch.

Parameters

model (Word2Vec or subclass) – Current model.

on_epoch_begin(model)

Method called at the start of each epoch.

Parameters

model (Word2Vec or subclass) – Current model.

on_epoch_end(model)

Method called at the end of each epoch.

Parameters

model (Word2Vec or subclass) – Current model.

on_train_begin(model)

Method called at the start of the training process.

Parameters

model (Word2Vec or subclass) – Current model.

on_train_end(model)

Method called at the end of the training process.

Parameters

model (Word2Vec or subclass) – Current model.

class gensim.models.callbacks.CoherenceMetric(corpus=None, texts=None, dictionary=None, coherence=None, window_size=None, topn=10, logger=None, viz_env=None, title=None)

Bases: gensim.models.callbacks.Metric

Metric class for coherence evaluation.

See also

CoherenceModel

Parameters
  • corpus ({iterable of list of (int, float), scipy.sparse.csc}, optional) – Stream of document vectors or sparse matrix of shape (num_documents, num_terms).

  • texts (list of char (str of length 1), optional) – Tokenized texts needed for coherence models that use sliding window based probability estimator.

  • dictionary (Dictionary, optional) – Gensim dictionary mapping from integer IDs to words, needed to create corpus. If model.id2word is present, this is not needed. If both are provided, dictionary will be used.

  • coherence ({'u_mass', 'c_v', 'c_uci', 'c_npmi'}, optional) – Coherence measure to be used. ‘c_uci’ is also known as ‘c_pmi’ in the literature. For ‘u_mass’, the corpus MUST be provided. If texts is provided, it will be converted to corpus using the dictionary. For ‘c_v’, ‘c_uci’ and ‘c_npmi’, texts MUST be provided. Corpus is not needed.

  • window_size (int, optional) –

    Size of the window to be used for coherence measures using boolean sliding window as their probability estimator. For ‘u_mass’ this doesn’t matter. If ‘None’, the default window sizes are used which are:

    • c_v - 110

    • c_uci - 10

    • c_npmi - 10

  • topn (int, optional) – Number of top words to be extracted from each topic.

  • logger ({'shell', 'visdom'}, optional) – Monitor training process using one of the available methods. ‘shell’ will print the coherence value in the active shell, while ‘visdom’ will visualize the coherence value with increasing epochs using the Visdom visualization framework.

  • viz_env (object, optional) – Visdom environment to use for plotting the graph. Unused.

  • title (str, optional) – Title of the graph plot in case logger == ‘visdom’. Unused.

get_value(**kwargs)

Get the coherence score.

Parameters

**kwargs

Key word arguments to override the object’s internal attributes. One of the following parameters are expected:

Returns

The coherence score.

Return type

float

set_parameters(**parameters)

Set the metric parameters.

Parameters

**parameters – Keyword arguments to override the object’s internal attributes.

class gensim.models.callbacks.ConvergenceMetric(distance='jaccard', num_words=100, n_ann_terms=10, diagonal=True, annotation=False, normed=True, logger=None, viz_env=None, title=None)

Bases: gensim.models.callbacks.Metric

Metric class for convergence evaluation.

Parameters
  • distance ({'kullback_leibler', 'hellinger', 'jaccard'}, optional) – Measure used to calculate difference between any topic pair.

  • num_words (int, optional) – The number of most relevant words used if distance == ‘jaccard’. Also used for annotating topics.

  • n_ann_terms (int, optional) – Max number of words in intersection/symmetric difference between topics. Used for annotation.

  • diagonal (bool, optional) – Whether we need the difference between identical topics (the diagonal of the difference matrix).

  • annotation (bool, optional) – Whether the intersection or difference of words between two topics should be returned.

  • normed (bool, optional) – Whether the matrix should be normalized or not.

  • logger ({'shell', 'visdom'}, optional) – Monitor training process using one of the available methods. ‘shell’ will print the coherence value in the active shell, while ‘visdom’ will visualize the coherence value with increasing epochs using the Visdom visualization framework.

  • viz_env (object, optional) – Visdom environment to use for plotting the graph. Unused.

  • title (str, optional) – Title of the graph plot in case logger == ‘visdom’. Unused.

get_value(**kwargs)

Get the sum of each element in the difference matrix between each pair of topics in two topic models.

A small difference between the partially trained models produced by subsequent training iterations can indicate that the model has stopped significantly improving and has therefore converged to a local or global optimum.

Parameters

**kwargs – Key word arguments to override the object’s internal attributes. Two models of type LdaModel or its wrappers are expected using the keys model and other_model.

Returns

The sum of the difference matrix between two trained topic models (usually the same model after two subsequent training iterations).

Return type

float

set_parameters(**parameters)

Set the metric parameters.

Parameters

**parameters – Keyword arguments to override the object’s internal attributes.

class gensim.models.callbacks.DiffMetric(distance='jaccard', num_words=100, n_ann_terms=10, diagonal=True, annotation=False, normed=True, logger=None, viz_env=None, title=None)

Bases: gensim.models.callbacks.Metric

Metric class for topic difference evaluation.

Parameters
  • distance ({'kullback_leibler', 'hellinger', 'jaccard'}, optional) – Measure used to calculate difference between any topic pair.

  • num_words (int, optional) – The number of most relevant words used if distance == ‘jaccard’. Also used for annotating topics.

  • n_ann_terms (int, optional) – Max number of words in intersection/symmetric difference between topics. Used for annotation.

  • diagonal (bool, optional) – Whether we need the difference between identical topics (the diagonal of the difference matrix).

  • annotation (bool, optional) – Whether the intersection or difference of words between two topics should be returned.

  • normed (bool, optional) – Whether the matrix should be normalized or not.

  • logger ({'shell', 'visdom'}, optional) – Monitor training process using one of the available methods. ‘shell’ will print the coherence value in the active shell, while ‘visdom’ will visualize the coherence value with increasing epochs using the Visdom visualization framework.

  • viz_env (object, optional) – Visdom environment to use for plotting the graph. Unused.

  • title (str, optional) – Title of the graph plot in case logger == ‘visdom’. Unused.

get_value(**kwargs)

Get the difference between each pair of topics in two topic models.

Parameters

**kwargs – Key word arguments to override the object’s internal attributes. Two models of type ldamodelLdaModel or its wrappers are expected using the keys model and other_model.

Returns

  • np.ndarray of shape (model.num_topics, other_model.num_topics) – Matrix of differences between each pair of topics.

  • np.ndarray of shape (model.num_topics, other_model.num_topics, 2), optional – Annotation matrix where for each pair we include the word from the intersection of the two topics, and the word from the symmetric difference of the two topics. Only included if annotation == True.

set_parameters(**parameters)

Set the metric parameters.

Parameters

**parameters – Keyword arguments to override the object’s internal attributes.

class gensim.models.callbacks.Metric

Bases: object

Base Metric class for topic model evaluation metrics.

Concrete implementations include:

get_value()

Get the metric’s value at this point in time.

Warning

The user must provide a concrete implementation for this method for every subclass of this class.

Returns

The metric’s type depends on what exactly it measures. In the simplest case it might be a real number corresponding to an error estimate. It could however be anything else that is useful to report or visualize.

Return type

object

set_parameters(**parameters)

Set the metric parameters.

Parameters

**parameters – Keyword arguments to override the object’s internal attributes.

class gensim.models.callbacks.PerplexityMetric(corpus=None, logger=None, viz_env=None, title=None)

Bases: gensim.models.callbacks.Metric

Metric class for perplexity evaluation.

Parameters
  • corpus ({iterable of list of (int, float), scipy.sparse.csc}, optional) – Stream of document vectors or sparse matrix of shape (num_documents, num_terms).

  • logger ({'shell', 'visdom'}, optional) – Monitor training process using one of the available methods. ‘shell’ will print the perplexity value in the active shell, while ‘visdom’ will visualize the coherence value with increasing epochs using the Visdom visualization framework.

  • viz_env (object, optional) – Visdom environment to use for plotting the graph. Unused.

  • title (str, optional) – Title of the graph plot in case logger == ‘visdom’. Unused.

get_value(**kwargs)

Get the coherence score.

Parameters

**kwargs – Key word arguments to override the object’s internal attributes. A trained topic model is expected using the ‘model’ key. This can be of type LdaModel, or one of its wrappers, such as LdaMallet or LdaVowpalWabbit.

Returns

The perplexity score.

Return type

float

set_parameters(**parameters)

Set the metric parameters.

Parameters

**parameters – Keyword arguments to override the object’s internal attributes.