search

GuHongyang / VaDE-pytorch

the reproduce of Variational Deep Embedding : A Generative Approach to Clustering Requirements by pytorch
121 stars 25 forks source link

hyper-parameters #1

Open MiguelMonteiro opened 4 years ago

MiguelMonteiro commented 4 years ago

Hello,

Thanks for sharing this code. I was wondering if you could also share the hyper-parametrization you have used in order to obtain the training curve on the front page: 94% cluster accuracy after 300 epochs. Running the code out of the box it seems to plateau at around 80%

Thank you,

Best regards,

Miguel

Jillian555 commented 4 years ago

I have the same problem!!!

YUMI66666 commented 4 years ago

I just run code it can get the 94%+ accuracy..

zhangzs-pku commented 4 years ago

I meet the same problem, and just setting --hid_dim 50 can obtain 94%+ acc

goodman1204 commented 4 years ago

I meet the same problem, and just setting --hid_dim 50 can obtain 94%+ acc

I set the hid_dim as 50, get worse acc, any ideas?

wzm2256 commented 4 years ago

The variance of the method is quite high. There is no way to avoid this. Try to repeat the experiment at least 10 times.

GuHongyang commented 4 years ago

This method relies on the GMM to initialize the network, and you can get the final result by trying a few more times!

wzm2256 commented 4 years ago

@GuHongyang Yes, but that is unfair. Because in practice you simply do not have any label to evaluate your results.

GuHongyang commented 4 years ago

Yes. The core problem is still the algorithm itself. You can divide the training data into a part of the data as validation data, and use this data to find better GMM initialization parameters.

goodman1204 commented 4 years ago

The variance of the method is quite high. There is no way to avoid this. Try to repeat the experiment at least 10 times.

Thank for your comments. I am wondering if the accuracy will be infected if I don't use GPU ( I removed the cuda() part).

RenShijie commented 3 years ago

Yes. The core problem is still the algorithm itself. You can divide the training data into a part of the data as validation data, and use this data to find better GMM initialization parameters.

Thanks for the advice. I have the same problem, I wonder if the network is correctly initialized. I ran the code for about 10 times with PyTorch 1.7, but it can never achieve 82%+. Is there any difference between different PyTorch versions in network initialization? Besides, can you share the final code in this repository?