search

pathak22 / context-encoder

[CVPR 2016] Unsupervised Feature Learning by Image Inpainting using GANs
https://www.cs.cmu.edu/~dpathak/context_encoder/
Other
876 stars 204 forks source link

Random missing areas get bad effect #22

Open windSong1 opened 6 years ago

windSong1 commented 6 years ago

hello! Thanks for your brilliant paper and code first.

I use the 'train_random.lua' script to train the net to complete an image with random loss area,I train it in a GPU of GTX 1080, I found the train time is so long if I loop 500 times, almost 150 hours, so I only loop 80 times.the effect in visulization web is good,but when I test it in validation data, the result is so bad.

Followings are some results of my training in random areas,the effet is bad screenshot from 2017-11-19 17-28-00

Followings are resuls from weight in center missing that you offered for me, the result is good. screenshot from 2017-11-19 17-33-34

is it the reason that my epoch loops is not enough? any advices will be appreciated.

windSong1 commented 6 years ago

I got it, I trained it with fineSize of 64x64, but I test it with fineSize of 128x128, but when I use fineSize of 64x64 to test, the result is still not good,I only trained it 80 epoch because of the time.

screenshot from 2017-11-21 10-59-31

maryam089 commented 6 years ago

@windSong1 you mean results are good At training with fine size 64x64 and testing at 128x128 ?

windSong1 commented 6 years ago

@maryam089 no, I got a mistake at first, as you can see, the result is not good if the train size and test size are not matching. but when I match them all, the resul is still not good, see the last picture. I guess it is the reson that my training epoches are not enough,only 80 epoches.

windSong1 commented 6 years ago

@maryam089 I trained it on flying-chair dataset,not ImageNet,do you still want it?

micklexqg commented 6 years ago

@windSong1 , I notice that the random position is also fixed from your result image. if i want to inpaint other areas or anywhere. how should i do ?

NerminSalem commented 5 years ago

did u tried to activate adv loss?