BERTopic Embeddings ValueError when transform a new text

Question

I have created embeddings using SentenceTransformer and trained a BERTopic model on those embeddings.

sentence_model = SentenceTransformer("all-MiniLM-L6-v2")
embeddings = sentence_model.encode(training_docs, show_progress_bar=True)
topic_model = BERTopic().fit_transform(training_docs, embeddings)
topic_model.reduce_topics(training_docs, nr_topics=5)

I have then saved the embeddings using pickle and topic_model using topic_model.save(). I can also load them both but when I try to use it on a new text such as:

with open('embeddings.pickle', 'rb') as pkl:
    embeddings = pickle.load(pkl)

topic_model = BERTopic.load('mybertopic')

sentence = 'I have found my car.'

topics, probs = topic_model.transform(sentence, embeddings)

I get the following error:

ValueError: Make sure that the embeddings are a numpy array with shape: (len(docs), vector_dim) where vector_dim is the dimensionality of the vector embeddings. 

The embeddings are a numpy array. How do I solve this?

Answer 1

Okay I solved it. I have to encode my text using the same SentenceTransformer and not use the entire embeddings in transform method.

embeddings = sentence_model.encode(sentence)
topics, probs = topic_model.transform(sentence, embeddings)
print(topics)
[-1]