search

taoxugit / AttnGAN

MIT License
1.33k stars 415 forks source link

The inception score on coco dataset #46

Closed liushanyuan18 closed 4 years ago

liushanyuan18 commented 5 years ago

Hi,in my experiment,the inception score of the pretrained model on coco dataset is 16.16.Do you know why the IS is not stable?How can make the IS to 25.89?

mkumar10 commented 5 years ago

I think your evaluation is flawed. I had that issue. Make sure your images are in RGB format and not BGR if you use opencv

liushanyuan18 commented 5 years ago

I think your evaluation is flawed. I had that issue. Make sure your images are in RGB format and not BGR if you use opencv

Thank you so much!You are right. By the way,I have a question in CA(Conditioning Augmentation).We convert embedding ϕt to ĉ0,and use the KL loss between ĉ0 and N~(0,1) to make ĉ0 close to N~(0,1).Why the KL loss is necessary?If ĉ0 is far away from N~(0,1),I think this is OK and has no bad,but when I abandon KL loss,the IS and FID is not good as early.Do you know why?Thanks!

ShihuaHuang95 commented 5 years ago

@liushanyuan18 @mkumar10 How about your scores? In my experiment, the IS of the pre-trained model on coco dataset is 23.51, it seems far from the 25.89.

liushanyuan18 commented 5 years ago

@liushanyuan18 @mkumar10 How about your scores? In my experiment, the IS of the pre-trained model on coco dataset is 23.51, it seems far from the 25.89.

We get 23.74 on coco dataset.It is similar to your score.