Hi,
Thank you for sharing the QIML code. I am trying to reproduce the method and would like to clarify the QPrior representation used in predictor regularization.
In QPRIOR_kf_final.py, the QPrior seems to be trained as a binned velocity/PDF-style distribution, using bin edges / bin centers and a target one-hot histogram:
https://github.com/UCL-CCS/QIML/blob/main/script/QPRIOR_kf_final.py#L45-L47
https://github.com/UCL-CCS/QIML/blob/main/script/QPRIOR_kf_final.py#L75-L97
However, in train_q_kf.py, the rollout distribution appears to be constructed by flattening the predicted velocity magnitude and applying softmax, before comparing it with QG_prior:
https://github.com/UCL-CCS/QIML/blob/main/script/train_q_kf.py#L172-L188
Could you please clarify which representation was intended for the experiments in the paper?
Thank you!
Hi,
Thank you for sharing the QIML code. I am trying to reproduce the method and would like to clarify the QPrior representation used in predictor regularization.
In QPRIOR_kf_final.py, the QPrior seems to be trained as a binned velocity/PDF-style distribution, using bin edges / bin centers and a target one-hot histogram:
https://github.com/UCL-CCS/QIML/blob/main/script/QPRIOR_kf_final.py#L45-L47
https://github.com/UCL-CCS/QIML/blob/main/script/QPRIOR_kf_final.py#L75-L97
However, in train_q_kf.py, the rollout distribution appears to be constructed by flattening the predicted velocity magnitude and applying softmax, before comparing it with QG_prior:
https://github.com/UCL-CCS/QIML/blob/main/script/train_q_kf.py#L172-L188
Could you please clarify which representation was intended for the experiments in the paper?
Thank you!