search

Jasonkks / PTTR

Pytorch Implementation of PTTR: Relational 3D Point Cloud Object Tracking with Transformer
Apache License 2.0
121 stars 12 forks source link

Unable to reproduce #3

Closed botaoye closed 2 years ago

botaoye commented 2 years ago

Hi, thanks for your work. I try to train the Kitti Cyclist model with your code using the default training setting (160 epochs with batch size 64), but the performance of the last epoch (58.0/78.9) is largely behind the result in your paper (65.1 / 90.5). So I wonder if there are any tricks or common practices to picking the model (e.g. finding the model with the best performance rather than the model of the last epoch)?

Jasonkks commented 2 years ago

Hi, thanks for your interest. I reran the training and testing, but it seems that I do not encounter the issue you mentioned. There might be some fluctuations in performance, but the gap you mentioned is not common from my experience. I have attached the train/test log and the checkpoint for your reference at link.