Hello,
I just completed GAN C1 W4 lab and wanted to ask how the lab’s implementation relates to the implementation in the Interpreting the Latent Space of GANs… paper. My understanding is that each feature (semantic attribute, like “smiling”) is a vector in the latent space (the space of the 64 element noise vectors), and that the set of 40 features were trained using the feature labels in the training set.
My questions are:
- Is the technique we used to train to find a single latent space feature (i.e., semantic attribute, e.g., “smiling”) similar to the SVM method the paper uses?
We train five independent linear SVMs on pose, smile, age, gender, eyeglasses, and then evaluate them on the validation set (6K samples with high confidence level on attribute scores) as well as the entire set (480K random samples).
- Is the technique we used to disentangle the feature we are interested in (e.g., “smiling”) from all of the other features the same as what the paper calls “Conditional Manipulation”
Taking the left sample in Fig.7 as an example, the results tend to become male when being edited to get old (first row). We fix this problem by subtracting its projection onto the gender direction (second row) from age direction, resulting in a new direction. In this way, we can make sure the gender component is barely affected when the sample is moved along the projected direction
Thanks!
Steve