Clarification on Model Update after Outlier Reduction with Images-Only

[owwdesilva]

Hi @MaartenGr, Hi All,

Clarification on an issue encountered with the model using an images-only training approach. Below is the relevant code snippet we used to train the model:

hdbscan_model = HDBSCAN(min_cluster_size=best_hyperparameters['min_cluster_size'], metric='euclidean', prediction_data=True)
vectorizer = CountVectorizer(stop_words="english", ngram_range=best_hyperparameters['n_gram_range'])

topic_model = BERTopic(
        embedding_model=MultiModalBackend, 
        representation_model=representation_model,
        umap_model=umap_model,
        hdbscan_model=hdbscan_model,
        vectorizer_model=vectorizer,
        ctfidf_model=ctfidf_model,
        calculate_probabilities= True,
        verbose=True,
        top_n_words=best_hyperparameters['top_n_words']
)

topics, probs = topic_model.fit_transform(documents=None, images=images, embeddings=image_embeddings)

We observed outliers in the output and proceeded with the outlier reduction method as follows:

new_topics = topic_model.reduce_outliers(None, topics, images=images, probabilities=probs, strategy="probabilities")

However, we encountered an issue when attempting to update the model, as the update_topics method requires documents to be passed:

topic_model.update_topics([], images=images, topics=new_topics)

Given that our model is trained using only images, providing documents in this context is not feasible. Could you please provide advice or guidance on how to address this issue?

[MaartenGr]

This is interesting as I haven't used this images-only procedure with .reduce_outliers and .update_topics. In all honesty, based on the code I'm seeing in .update_topics, I don't think this will work unfortunately. Instead, you could use manual topic modeling using the topics you have previously created.

[owwdesilva]

Thanks @MaartenGr for the insight. I'll try manual topic modeling with the existing topics.