你好

我阅读了https://maartengr.github.io/BERTopic/api/bertopic.html#bertopic._bertopic.BERTopic.transform,并从文档中的参数(描述为“单个文档或要预测的文档列表”)了解到，即使使用拟合模型进行预测，我也可以将文档列表或单个文档提交，仍然可以获得相同的结果。

我发现这并不是真的。我是不是忽略了什么？

以下是一个最小工作示例：

from bertopic import BERTopic
from sklearn.datasets import fetch_20newsgroups

docs = fetch_20newsgroups(subset='all')['data'][:200]
topic_model = BERTopic().fit(docs)
topics,_=topic_model.transform(docs)

import numpy as np
topics=np.array(topics)

#calling the model with a single document several times
import tqdm

topics_single = []
for doc in tqdm.tqdm(docs):
    topic, _ = topic_model.transform([doc])
    topics_single.append(topic[0])
topics_single = np.array(topics_single)

mask_identical = topics_single == topics
percentage_equal = 100 * np.sum(mask_identical) / len(mask_identical)
print(f"{percentage_equal=}%") #returns for example about 60%, but varying

#loop till finding a different entry
for i in range(len(docs)):
    print(
        i,
        topic_model.transform(docs[i])[0][0],
        topic_model.transform([docs[i]])[0][0],
        topics[i],
        topic_model.transform(docs)[0][i],
    )
    if topic_model.transform(docs[i])[0][0] != topic_model.transform(docs)[0][i]:
        print(f"Different outcome at iteration {i}")
        break

重复执行相同文档似乎没问题：

topics2,_=topic_model.transform(docs)
percentage_equal_executed_with_multiple_docs = 100*np.sum(np.array(topics2)==topics)/len(topics)
print(f"{percentage_equal_executed_with_multiple_docs=}%") #this gives 100%

提前感谢！
PS:
Python版本是3.10.12
已安装的包列表如下：