search

ZZUTK / Face-Aging-CAAE

Age Progression/Regression by Conditional Adversarial Autoencoder
636 stars 233 forks source link

Cannot reproduce results #27

Open geopapyrus opened 6 years ago

geopapyrus commented 6 years ago

Hi everyone!

We are getting strange results where everybody looks the same!

I've downloaded the UTKFace.tar.gz dataset as described and trained with the command "python main.py" and all default parameters.

Attached is a gif showing all 50 epochs during training.

ezgif-1-25017abe1f

Then we tried using some images from the training set to see how well the test output would be.

And this was the result:

Input:

input

Female output.

test_as_female

Male output.

test_as_male

Does anybody know what we need to change in order to get good results?

thank you

liyin2015 commented 6 years ago

I think the problem is the oversmoothing because they imposed a uniform distribution on their latent variable z. You can get way better result if you remove the discriminator on z, or try to put normal distribution to replace the uniform distribution

praveenkumarchandaliya commented 6 years ago

I am using same data set with Python3.5 and Tensorflow 1.7 , Teslak40 but I am getting blank results. can you suggest me which type of problem occurred or which thing use change in your code?