guhongyang / vade-pytorch Goto Github PK

Thank you for sharing your code.

In the paper, pi_.sum() = 1. I found that vade.pi_.sum() > 1 in this code. Why?

Hi @GuHongyang ,
Thanks for the reimplementation!
they mention in the paper that As for the generative process in Section 3.1, the multivariate Bernoulli distribution is used for MNIST dataset, and the multivariate Gaussian distribution is used for the others . In your code, you've implemented the multivariate Gaussian distribution right?
Sincerely,
Thibault

After pre training, what is the prediction accuracy? My acc is only 0.7%. why?

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

Thanks for your pytorch code. I know the pi_c is obtained from GMM . however, when I get a pi_c, the pi_c[0] represents the label 0's prior probability from GMM, how I know the label 0 from GMM equals to the label 0 from the final clustering result. for example ,a tensor pi_c=[0.1,0.5,0.4],which represents the prior probability of label 0 is 0.1,label 1 is 0.5,label 2 is 0.4,does is it align to the label of q(c|x).I don't know if I explain clearly. In short, because this is a unsupervise model,so maybe the GMM's label is not aligned to the q(c|x)'label.

Any one knows this bug? I just run the original code and show the following bug.

Traceback (most recent call last):
File "main.py", line 50, in
vade.module.pre_train(DL,pre_epoch=50)
File "/home/shuiqiao/anaconda3/lib/python3.7/site-packages/torch/nn/modules/module.py", line 594, in getattr
type(self).name, name))
AttributeError: 'VaDE' object has no attribute 'module'
'''

Hi, in your model.py,

@staticmethod
    def gaussian_pdf_log(x,mu,log_sigma2):
        return -0.5*(torch.sum(np.log(np.pi*2)+log_sigma2+(x-mu).pow(2)/torch.exp(log_sigma2),1))

I wonder why you add 1? I think in Gaussian distribution, the pdf is

-0.5*(np.log(np.pi*2)+log_sigma2+(x-mu).pow(2)/torch.exp(log_sigma2))

Hello, thank you for sharing the code.

In the original paper, pi are parameters for categorical variable, so it must sum to one.
However, in this code, when I print out the summation of pi, it exceeds the value one.
Shouldn't there be some kind of softmax or noramlization to prevent this? (SoftMax seems to degrade the performance though...)

I have solved this question, sorry about disturbing you!