Hifill-tensorflow
Hifill-tensorflow copied to clipboard
duxingren14
the result is not good
I retrain the model for about 800000 pics of image in Celeba-HQ, the final result is not good, even bad comparing to other gan based model. It just a random inpainting. I think may you result is over-fitting in you small dataset.
Which model? the one trained on places2?
On Tue, Oct 25, 2022 at 10:21 AM Shawn-Yu-1 @.***> wrote:
I retrain the model for about 800000 pics of image in Celeba-HQ, the final result is not good, even bad comparing to other gan based model. It just a random inpainting. I think may you result is over-fitting in you small dataset.
— Reply to this email directly, view it on GitHub https://github.com/duxingren14/Hifill-tensorflow/issues/9, or unsubscribe https://github.com/notifications/unsubscribe-auth/AATTU666LIDGUA2VJIG2T6DWE472XANCNFSM6AAAAAARNQZLWI . You are receiving this because you are subscribed to this thread.Message ID: @.***>
I said trained on celeba-hq，why you said places2？ I only changed batchsize and dataset，but the result is bad. all patchs are inpainting with the same color.
No idea what happens to your training. I think it is unreasonable to generate "random" results if the training is properly performed. Assume that a model could be overfit to a small dataset, then a set of facial images should be easier to overfit.
My personal advice is, make sure the facial images you used for training are well aligned and correctly preprocessed. I guess CelebA-HQ only contains 30k images, where do the 800k images come from? Plus, make sure the losses are properly set up and mode collapse does not happen.
BTW, the official model is trained on a subset of places2 and never met the test images presented on the paper.
On Wed, Oct 26, 2022 at 8:37 AM Shawn-Yu-1 @.***> wrote:
I said trained on celeba-hq，why you said places2？ I only changed batchsize and dataset，but the result is bad. all patchs are inpainting with the same color.
— Reply to this email directly, view it on GitHub https://github.com/duxingren14/Hifill-tensorflow/issues/9#issuecomment-1291284167, or unsubscribe https://github.com/notifications/unsubscribe-auth/AATTU63BRBMMY5DKOIUHOQDWFB4NRANCNFSM6AAAAAARNQZLWI . You are receiving this because you commented.Message ID: @.***>
you may misunderstand my words. I said "train 800000 pics of image" mean use the batch size of 8 run about 100000 iters. I just want to verify the result in the paper, and it runs very slow. yesterday I ran the deepfillv2 , this is faster. I use the batchsize of 16 run 100000 iters just cost the half time of the hifill. so, I may don't have time to figure out the reason of the bad result, just let the issue over.