thuiar / deepaligned-clustering Goto Github PK

doest the DAC means Deep Adaptive Clustering?

paper source like this url: https://openaccess.thecvf.com/content_ICCV_2017/papers/Chang_Deep_Adaptive_Image_ICCV_2017_paper.pdf

Hi, @HanleiZhang
The paper - 'Discovering New Intents with Deep Aligned Clustering' is quite interesting. For some experiments, I need to know the actual numbers in the Fig. 3, 4, 5 and 6. Can you please provide the CSV files which are used for creating the figures? That would be great. Thanks

请问一下，在训练完之后，得到了bert分类模型，类别数是self.num_labels。
在测试时，我看代码中是这样处理：先用bert提取测试集的embedding，然后用kmeans聚类，类数是self.num_labels，得到聚类结果以后去和真实标签计算最优匹配。
那么请问测试时为什么不直接对测试集分类呢？直接分类应该也可以得到对应的label

作者您好，最近在做新意图发现的工作，在复线这篇工作时，我发现直接在预训练模型的基础上+Kmeans在ACC和ARI两个数据上可以达到比预训练+DAC基本上高2-3个点，MNI基本上差不多。目前我也不确定这是否是正常的现象还是我使用有误，如果您之前有类似的经验，希望可以解答一下这个现象，十分感谢！
我的实验设置是banking+0.75（known）+0.1（labeled），K设置为77，对Kmeans的实现方法为直接在ModelManager的train之前调用test函数得到Kmeans结果。

I'm interested in your work. Please, how to distinguish new intents and old intents(known intents) in the test phase?

Hi,
I notice that this paper uses CLINC and BANKING dataset. Your previous work (Discovering new intents via constrained deep adaptive Clustering with Cluster Refinement) uses SNIPS, DBPedia, StackOverflow dataset. It seems that this two studies study the same task? And what is the benchmark dataset which used in the future from your perspective?

Hi @HanleiZhang
In file : https://github.com/thuiar/DeepAligned-Clustering/blob/main/results/k_results.csv
The second CSV file (Varying value of K) that you have shared, contains ACC wrt different approaches but the predicted value of K is not there. Did you run it for 10 seeds? And if so, then there will be 10 different values of K predicted. Can you please also share the Kpred value for each approach. Especially, Deep Aligned.

感谢作者大大开源您的代码，我最近看了您的CDAC+和这篇模型，都发现您对BERT的前11层参数都采取了冻结的做法，我想问一下如果不冻结前11层参数，效果会是怎样的？还是说这样做是考虑到训练显存的问题，冻结前11层的参数可以在训练时设置更大的batch_size? 希望得到您的解答