huan ling
Interesting finding! In my experience, the editing vector is not that view conditioned. But you may need to adjust the scale.
In my case, I got 31G for training data.
did you change the line "self.optimizer = tf.train.AdamOptimizer" in __init__ ?
It's by HopeNet[1] if I understand correctly. [1]: _Fine-Grained Head Pose Estimation Without Keypoints,_ **Ruiz, Nataniel and Chong, Eunji and Rehg, James M.**