Running inference on custom skeletons

[icedwater]

Hello Jihoon, thanks for sharing this code. I've adapted it to take a different model for a separate project by creating new variables in cmib/model/skeleton.py, but I can't seem to produce meaningful inferences.

I have a few questions about the inference process that I hope you can take some time to help me with.

1. From your experience, is there a recommended minimum input training length? I see LAFAN1 has sequences from 3k frames to 9k frames long, but I currently have sequences under 1k.
2. Can I confirm that from_idx and target_idx are loaded from the checkpoint and that by default, there is no way to specify them at inference time?
3. Is there a way to generate more frames over the given inference interval? e.g. between frame 10 and 50, I would like to make a longer sequence that is about 100 frames, to support different playback rates.
4. For test_idx, what do the values 950, 1140, 2100 correspond to? I thought they were frame numbers, but I can't seem to give numbers more than the test classes that I have.

Thanks!

[icedwater]

Hello @childtoy if you could kindly provide some hints as well in case Jihoon is too busy, I would be very grateful. Thank you!