search

mohamedhassanmus / SAMP

Stochastic Scene-Aware Motion Prediction https://samp.is.tue.mpg.de/
133 stars 15 forks source link

About the training details in index selection #4

Closed SternenHimmel0299 closed 2 years ago

SternenHimmel0299 commented 2 years ago

Thanks a lot for your great work! One question about the training code in MotionNet_train.py, line 35-40:

for i in range(args.L): optimizer.zero_grad() p = y[:, i, :] p_prev = x1[:, i, :] I = x2[:, i, :] So i am assuming that the rollout length is the sequence length which is determined to be 60 in your implemetation. I am wondering in the encoding stage, why the indices for both previous pose p_prev and next pose p are set to i. Does that mean the i-th pose in p_prev is the corresponding past pose for the i-th pose in p?

I would appreciate it very much if you could reply to my quesytion. Thank you!

mohamedhassanmus commented 2 years ago

Yes. It is exactly as you guessed. The i-th pose in p_prev is the corresponding past pose for the i-th pose in p.

SternenHimmel0299 commented 2 years ago

Thank you for the kind reply! That helped me a lot.