Problem with shape of Child model theta matrix

[NikolayPavlychev]

model = artm.ARTM(
            class_ids=modalities_list,
            num_topics=num_topics[1],
            scores=scores,
            topic_names=[
                "topic" + str(t) for t in range(num_topics[1])
            ],
            cache_theta=True,
        )
        model.num_document_passes = 1
        model.initialize(dictionary)


        model.fit_offline(batch_vectorizer=batch_vectorizer, num_collection_passes=1)

        num_topics_child = num_topics[2] # 30
        topic_names_child = ['topic_{}'.format(i) for i in range(num_topics_child)]

        child_model = artm.ARTM(
            seed=1,
            topic_names=topic_names_child,
            cache_theta=True,
            class_ids=modalities_list,
            parent_model=model,  # when specified get_theta has wrong number of columns
            parent_model_weight=2,
            dictionary=dictionary,
            num_document_passes=1,
            #theta_name='child_theta'
        )

      child_model.fit_offline(batch_vectorizer=batch_vectorizer, num_collection_passes=90)

for 41 documents, it supposed to be 30x41, but instead 30x (41 + 10 = 51) returned
initial column names in theta contain only 0 (zeroes). So it’s like 0 0 0 0 0 0 0 0 0 0 0 0, than goes real columns 1, 2, 3, … Those zeroes are appended to the end with increasing num_collection_passes. So with num_collection_passes=90, theta shape is 30 x 941.
print(child_model.get_theta())

BIGARTM documentation says that theta supposed to be number of topics X number of documents.**

[bt2901]

This is expected behaviour for hierarchical models. In that case, ARTM adds 'pseudodocuments' to the collection and tries to factorize this new collection. Each pseudodocument is related to some topic of the parent model (since the distributions p(subtopic|doc) and p(subtopic|parent_topic) are similar computationally). If you run child_model.get_parent_psi(), the Psi matrix you obtain should consist of the "extra" columns of Theta, precisely.

Also, for convergence/memory efficiency reasons the following patter is suggested:

model = artm.ARTM(
            ...
            # cache_theta=False,
            theta_columns_naming='title'
)
model.fit_offline(batch_vectorizer, ...)
theta = model.transform(batch_vectorizer)

The result is pandas DataFrame where the index is topic names, and columns are document names (without pseudodocuments).

You are absolutely correct that this could be made more clear in the documentation.