Daquisu
commented
3 years ago Hello,
In the CVPR paper they said:
We train an auxiliary attribute prediction model using the annotations from the CelebA dataset [26] with ResNet- 50 network [18]. This model is trained with multi-task losses to simultaneously predict smile, age, gender, eye- glasses, as well as the 5-point facial landmarks. Here, the facial landmarks will be used to compute yaw pose, which is also treated as a binary attribute (left or right) in further analysis. Besides the landmarks, all other attributes are learned as bi-classification problem with softmax cross- entropy loss, while landmarks are optimized with l2 regres- sion loss. As images produced by PGGAN and StyleGAN are with 1024×1024 resolution, we resize them to 224×224 before feeding them to the attribute model
Hi, I really thanks for your great research.
I want to try training boundaries so i need attribute score classifier to get the score. In some issues (ex, Latent code- attribute score pair #53), you will provide some classifier.
Is there any available classifier now that you provide? or just use ResNet-50, etc?
Thank you very much.