search

shepnerd / inpainting_gmcnn

Image Inpainting via Generative Multi-column Convolutional Neural Networks, NeurIPS2018
MIT License
428 stars 98 forks source link

training pytorch model on celebAHQ #20

Closed ljjcoder closed 5 years ago

ljjcoder commented 5 years ago

When I training the pytorch model on celebAHQ from the pretrain model you give, I watch the tensorboard result seems to be blurred like this. 1561199895(1) I train the model on 1080ti with batch size 4. could you give me some sugestion to get more real results?

shepnerd commented 5 years ago

I guess you may train the model with only the reconstruction loss. To train the model with all losses, you need set option pretrain_network as 0. Specifically, just as the follows: python train.py ... --pretrain_network 0 ....

ljjcoder commented 5 years ago

@shepnerd thanks for your reply, when I change set pretrain_network 0, it will produce perfect results training from pretrain model on rect mode. but when I train the stoke mode from scrach ,the reuslt seems to be very bad like this. Could you give me some sugestion to get more real results and could you give the stoke pretrain model to me? thanks.

1561361548(1)

shepnerd commented 5 years ago
  1. pretrain the model first (--pretrain_network 1)
  2. finetune the model with all losses when the previous training stage converges (--pretrain_network 0 --load_model_dir ...)

We do not have the pretrained model on celeba-hq using strokes with pytorch implementation. We may add it in the future. Maybe you could try our tensorflow version.

ljjcoder commented 5 years ago

@shepnerd I find the places2 model you give is trained on stroke, could you give me the model trained on rect?

shepnerd commented 5 years ago

We may release places2 model trained with rectangle masks in the future. Currently, you could just finetine model trained with strokes using rectangle masks.

ljjcoder notifications@github.com 于2019年6月27日周四 下午5:02写道：

@shepnerd https://github.com/shepnerd I find the places2 model you give is trained on stroke, could you give me the model trained on rect?

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/shepnerd/inpainting_gmcnn/issues/20?email_source=notifications&email_token=ABGU6CGAVFWEVN3MJDCRS2LP4R62RA5CNFSM4H2WRNFKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGODYWOWTQ#issuecomment-506260302, or mute the thread https://github.com/notifications/unsubscribe-auth/ABGU6CBMSCM5664W4A3QRTTP4R62RANCNFSM4H2WRNFA .

cmyyy commented 5 years ago

Hello, @ljjcoder, i met some problem when using tensorboard to do visualization. Could you share the envs of visualization, like the versions of tensorboard, tensorboardX, tensorflow etc.. Thanks a lot.