Questions about "DeepPhase: Periodic Autoencoders for learning motion phase manifolds"

[wengn]

Hi, Sebastian

How are you? We are following closely with your research work on applying Deep Learning onto character animation, and I want to say they are great work! We are reading your Siggraph 2022 paper "DeepPhase: Periodic Autoencoders for Learning Motion Phase Manifolds" and trying to reproduce the work, but got stuck on some questions. I am wondering if you could help me with these detailed questions.

1. What's the kernel-size of the convolultional layer?
2. What method did you use to initialize the weights?
3. What are the validation/test loss you achieved after you finished training?
4. If I change the kernal size, there are quite a few occations that loss became nan，do you know what could be the reason for this?
5. In the paper, does every channel connect to a unique fully connected layer? What's the activation function of the fully connected layer?
6. Does the FFT layer has weights to learn as well?
7. The sampling time for a time window is 2 second, correct? Also the T in "f" in formula (3) is also 2 seconds, right?

We used your dataset from paper "Neural State Machine for Character-Scene Interactions",but the lowest loss we could get is 0.2. We think it is too high and don't find a way to reduce it. Can you shed some light on this?

Avoid 18863（5.24min）
Carry 53094（14.75min）
Crouch 7659 （2.13min）
Door 58479（16.24min）
Jump 4511 （1.25min）
Loco 59859（16.63min）
Sit 199472（55.41min）
total: 401937 (111min)

Thanks a lot!