search

JiahuiYu / generative_inpainting

DeepFill v1/v2 with Contextual Attention and Gated Convolution, CVPR 2018, and ICCV 2019 Oral
http://jiahuiyu.com/deepfill/
Other
3.27k stars 788 forks source link

irregular mask train the model? #89

Closed Jason-xin closed 6 years ago

Jason-xin commented 6 years ago

Hello, sorry to bother you... You said it can not solve irregular mask in the version(#77)，you v2 paper is difficult for me... So, if I add irregular mask to your v1 model when training, it may be works well or not ?

What's more, when training Places2 model, all High-resolution images were used? It has 105G ... If I want to train my own dataset with the given Places2 pre-trained model given, how many images and size(256*256 mentioned in paper) are required? image

JiahuiYu commented 6 years ago

I use the high-resolution images, which is important. During training, the code will random crop a 256x256 patch in high-resolutions images.

Adding irregular mask to DeepFill v1 training will definitely improves the results. But it is not a fundamental solution for free-form masks. Thus I would still suggest to understand DeepFill V2, especially the gated convolution. If you still have questions after carefully reading the paper, just ask them here and I will try my best to help.

If you want to train your own dataset, more images will be better. For all experiments, at least I use 30,000 images (CelebA-HQ model). You could have a reference at issue #88. They have fairly good results in most cases.