search

qiaott / MirrorGAN

Pytorch implementation of MirrorGAN
129 stars 40 forks source link

re-trained model cannot get the IS score as the thesis provides #9

Open lovelychao opened 4 years ago

lovelychao commented 4 years ago

hello， I retrained the model as your thesis say， but the best IS score I get is only 3.89， is there any trick in training the model？ or can you provide your model?

MartinPernus commented 4 years ago

Same problem here, following the instructions, the final inception score on CUB is only 4.00

EDIT: By evaluating many trained models, one of them achieved IS of 4.26. I would be grateful if anyone has tips/tricks for improving this.

Kerry-zzx commented 4 years ago

there is no pretrained model!

WhereISmyroad commented 3 years ago

Hello sir I can train MirrorGAN by myself, but I need (image caption/text encoder) pretrained model( coco/CUB), too. The model （text encode/image caption）provided by K, I get 4.41 (batch size=12, batch num:3 , G&D LR:0.002)... I just want to do a comparative experiment, Can anyone help me？

YangShengLiQQ commented 3 years ago

Hello sir I can train MirrorGAN by myself, but I need (image caption/text encoder) pretrained model( coco/CUB), too. The model （text encode/image caption）provided by K, I get 4.41 (batch size=12, batch num:3 , G&D LR:0.002)... I just want to do a comparative experiment, Can anyone help me？

Hello @WhereISmyroad how did you get 4.41? is your code of inception score from https://github.com/hanzhanggit/StackGAN-inception-model ?

WhereISmyroad commented 3 years ago

Hello sir I can train MirrorGAN by myself, but I need (image caption/text encoder) pretrained model( coco/CUB), too. The model （text encode/image caption）provided by K, I get 4.41 (batch size=12, batch num:3 , G&D LR:0.002)... I just want to do a comparative experiment, Can anyone help me？

Hello @WhereISmyroad how did you get 4.41? is your code of inception score from https://github.com/hanzhanggit/StackGAN-inception-model ?

Yes, you are the same problem?

YangShengLiQQ commented 3 years ago

@WhereISmyroad Yes, I only got 4.0 at epoch 250, and 3.7 at epoch 650. I guess my code of inception score is wrong, because I use netG provided by author to generate about 29000 images. Then I use the code to evaluate them, but the score I got is only 2.5 Did you use the config file provided by author? How many epochs do you train? Could you share your config file with me? Thanks!

WhereISmyroad commented 3 years ago

@WhereISmyroad Yes, I only got 4.0 at epoch 250, and 3.7 at epoch 650. I guess my code of inception score is wrong, because I use netG provided by author to generate about 29000 images. Then I use the code to evaluate them, but the score I got is only 2.5 Did you use the config file provided by author? How many epochs do you train? Could you share your config file with me? Thanks!

I use the config file provided by this project, epoch=600 How you get the pretrained model? Its author didn't provide it

YangShengLiQQ commented 3 years ago

@WhereISmyroad Yes, I only got 4.0 at epoch 250, and 3.7 at epoch 650. I guess my code of inception score is wrong, because I use netG provided by author to generate about 29000 images. Then I use the code to evaluate them, but the score I got is only 2.5 Did you use the config file provided by author? How many epochs do you train? Could you share your config file with me? Thanks!

I use the config file provided by this project, epoch=600 How you get the pretrained model? Its author didn't provide it

The path of netG is \data\output\bird\Model, where the data is provided by author. https://drive.google.com/file/d/1CuW5ognTSkNbyx9TWoUFrgwqxZNk1cl0/view I don't know if it is pretrained model, because the score is only 2.5.

WhereISmyroad commented 3 years ago

@WhereISmyroad Yes, I only got 4.0 at epoch 250, and 3.7 at epoch 650. I guess my code of inception score is wrong, because I use netG provided by author to generate about 29000 images. Then I use the code to evaluate them, but the score I got is only 2.5 Did you use the config file provided by author? How many epochs do you train? Could you share your config file with me? Thanks!

I use the config file provided by this project, epoch=600 How you get the pretrained model? Its author didn't provide it

The path of netG is \data\output\bird\Model, where the data is provided by author. https://drive.google.com/file/d/1CuW5ognTSkNbyx9TWoUFrgwqxZNk1cl0/view I don't know if it is pretrained model, because the score is only 2.5.

I am not testing it yet, but I thank there is no problem with the IS Score code Do you have STREAM coco pretrained model? I train it by myself but get a bad score......