Cannot reproduce the demo from the blog post

[bamos]

Cannot reproduce the demo from the blog post

[bamos]

Hi, can you try lowering the learning rate and turning off momentum?

-Brandon.

[benstaf]

By dividing the learning rate by 100, and turning off momentum, the loss function decreases better. However, the result is still bad (especially the first picture, which is a selfie from myself, while the other 3 are famous politicians):

In your demo, what is the final loss?
I could not get less than 1500.

I ran the command:

./complete.py /home/ubuntu/dcgan-completion.tensorflow/static/aligned/rawfaces/* --lr 0.0001 --momentum 0.0 --nIter 1200 --outDir /home/ubuntu/dcgan-completion.tensorflow/static/outputimages

checkpoints here:

[bamos]

The last three there look much better than before, an adaptive technique would definitely help here. An adaptive one could potentially help the first image so you don't have to tune the learning rate. You could also try changing the lambda weight. I don't remember the final loss I got.

I'll close this issue for now but let me know if you notice anything else.

-Brandon.