Reproducing FID values of CIFAR and CELEBA after performing expost density estimation

[amruz]

Hi,
Really interesting work!
I am having trouble in reproducing the results for CIFAR/CELEBA images in table 1 for RAE versions. For MNIST I was able to reproduce the results, but for other two datasets I am getting high FID and nowhere around the reported values. Do you have any suggestions regarding what I did wrong? (I also trained from scratch and then tried with the uploaded autoencoder trained models (*.h5_best) as well)

Thanks in advance!

[amruz]

Hi,
Really interesting work!
I am having trouble in reproducing the results for CIFAR/CELEBA images in table 1 for RAE versions. For MNIST I was able to reproduce the results, but for other two datasets I am getting FID about 60 for CELEBA and 90 for CIFAR and is nowhere around the reported values. Do you have any suggestions regarding what I did wrong? (I also trained from scratch and then tried with the uploaded autoencoder trained models (*.h5_best) as well)

Thanks in advance!

[1994cxy]

Hi,
Could you please share your requirements.txt since I use the requirements in the project and got lots of error with six and keras backend. I use the Tensorflow==1.5.0 along with the Keras==2.3.1 and six=1.12.0