search

willylulu / celeba-hq-modified

Modified h5tool.py make user getting celeba-HQ easier
MIT License
121 stars 22 forks source link

what is the delta files for? #1

Closed micklexqg closed 5 years ago

micklexqg commented 5 years ago

Hi, thanks for your work, i wonder what is the delta files for? and have you get 1024x1024 images using h5tool.py?

willylulu commented 5 years ago

The celeba-hq images are not just downloaded by h5tool.py directly. First, it magnifies the images in celeba (that's why we need celeba dataset) to 1024x1024, then it uses delta files to unblur the images.

I create this repo because I fail to use the original h5tool.py to get celeba-hq. The original version needs lots of package and its compatibility is low.

The original h5tool.py will "create" the high-quality images and store in h5 file. I modify it to make it use fewer package and store the images directly.

micklexqg commented 5 years ago

@willylulu ， thank you, i run the code, but only get 30000 img files. i find that the given image_list.txt is not complete (only with 30000 rows). so how to get the complete txt files which is for 220599 image files?

willylulu commented 5 years ago

Please reference this markdown file https://github.com/aleju/papers/blob/master/neural-nets/Progressive_Growing_of_GANs.md In summary > how > CelebA-HQ section it says that : "They derive from CelebA images a new dataset containing 30k 1024x1024 images of celebrity faces."

I think it's only have 30000 images in celeba-hq. So training with celea-hq is quiet tough.

micklexqg commented 5 years ago

@willylulu ，好吧, that is not good。。。thank you！

Alalisho commented 4 years ago

hello, i really don't understand how can i convert my dataset celebA to celebA-HQ ,,,, can you help me please? what's the steps?