search

zhampel / clusterGAN

Pytorch Implementation of ClusterGAN (arXiv:1809.03627)
MIT License
95 stars 22 forks source link

not able to reproduce paper results #1

Open wangemm opened 4 years ago

wangemm commented 4 years ago

I have been running the code with the default params on MNIST, and using enc_zc_logits as the clustering results. And get the results: acc 0.56843 nmi 0.50773 ari 0.35441 fmi 0.44946

zhampel commented 4 years ago

@wangemm There are a number of parameters that can influence results. For example, the initialization of weights (line 51-71 in utils.py), and the width of the latent space distribution (line 81 in utils.py). The original author suggested for the latter a small value of like 0.15 for this, but I chose 0.75 during testing cause it avoided mode collapse on my system.

Furthermore, I did not include a classification model as in the author's original paper and repo. I assume you're using the architectures described in the paper...

wangemm commented 4 years ago

@zhampel Thanks! I`ll try.

TanmDL commented 4 years ago

@wangemm have you achieved the paper results?

wangemm commented 4 years ago

@TanmDL No,I can`t achieve the paper results.

YuxiaoWang-AI commented 4 years ago

code example 2 error，Please ask about the possible reasons

PFC-star commented 1 year ago

The paper proposes a modified backpropagation algorithm, but why is it not reflected in the code? image