goals.txt

—— Experimental Result Goals ——

Qualitative Results: Figure-7, Figure-8 on 256x256 resolution
Quantitative Results: Table-3 FID score comparison 3.25 on CELEBA and 3.10 on LSUN

Additional Results: FFHQ and 1024x1024 resolution experiments


—— version 1 submission ——
Unfortunately we could not generate results in our v1 submission since we had some problems
in our model implementations. We have detected several issues concerning the dimensions of signals
in our building blocks and they are fixed after v1.


—— version 2 submission ——
We have sucessfully trained our models and generated results. However, we have changed our goals due to training limitations that we have encountered during our experimentations.
We had a single gpu support unlikely to the setting in which the paper is originally experimented with
which corresponds to 8 32GB V100 GPUs and 32 batch size.
We couldn't fit more than 8 images for a batch. As a result we have decided to lower the resolution
for 256x256 images to 128x128 images in order to cut from training time and also have sufficiently many images for a batch.
In this setting, we were able to have batch sizes 4-8 and we have trained our models around 100k-300k of iterations
whereas the original paper's iterations are around 40-50 million.

Also our generator capacity is roughly the same with the original paper, whereas our discriminator capacity
is significantly less due to the limitational tradeoffs that we have made and also in the paper they didn't put
much emphasis on discriminator and we have adopted dcgan based a basic convolutional discriminator.

We have dropped our additional experimental goals on 1024x1024 resolution, as we can barely work on 128x128 resolution.
Additionally we have dropped FFHQ experiments and we have only focused on LSUN and CELEBA.

Our qualitative goals remains the same and we have generated images for both LSUN and CELEBA
corresponding to figures in the paper. We have mentioned to re-generate figures 7 and 8,
but we have genrated figure 12 and 13 which is a larger image grid of 5x4 serving the same purpose with more variety of images.