I have two question regarding hyperparameters for analyzing the model performance on p values.
If you have the hyperparameters to train p = 1, could you share the hyperparameters for the datasets (e.g. CelebA, Seg2Cat, edge2Car)?
The paper says
"Conversely, only sampling random poses (p = 1) gives the best image quality but suffers huge misalignment with input label maps."
I tried to train the proposed model with p = 1 several times with various parameters, but didn't succeed. I assume that it is because when p = 1, the model is not trained with CVC loss, and it makes the training procedure harder.
Additionally, I am wondering if the values on the chart from the Figure 9 are from CelebA dataset. Is it true?
Hello, Thank you for sharing your awesome work!
I have two question regarding hyperparameters for analyzing the model performance on p values.
If you have the hyperparameters to train p = 1, could you share the hyperparameters for the datasets (e.g. CelebA, Seg2Cat, edge2Car)?
The paper says "Conversely, only sampling random poses (p = 1) gives the best image quality but suffers huge misalignment with input label maps."
I tried to train the proposed model with p = 1 several times with various parameters, but didn't succeed. I assume that it is because when p = 1, the model is not trained with CVC loss, and it makes the training procedure harder.
Additionally, I am wondering if the values on the chart from the Figure 9 are from CelebA dataset. Is it true?
I greatly appreciate your work, thank you!