Open chenxiao18481132 opened 4 years ago
In my experience running this code (and even the original code released by the authors), the results are highly related to the initialization of the VaDE (i.e the pretrained autoencoder). Try re-running the code. Sometimes it will give you ~90% and sometimes ~80%. I did not perform a complete analysis, but I guess that if you train the autoencoder using a validation set and you save the best model within several epochs, you will reach better results.
I run your code in my computer, but I didn't get the good results as your figure (acc about 80%). Do I need to reset the hyper-parameters?